Scene Reconstruction and Visualization from Internet Photo Collections

Keith N. Snavely
A dissertation submitted in partial fulllment of
the requirements for the degree of
Doctor of Philosophy
University of Washington
2008
Program Authorized to Offer Degree: Computer Science & Engineering
University of Washington
Graduate School
This is to certify that I have examined this copy of a doctoral dissertation by
Keith N. Snavely
and have found that it is complete and satisfactory in all respects,
and that any and all revisions required by the nal
examining committee have been made.
Chair of the Supervisory Committee:
Steven M. Seitz
Reading Committee:
Steven M. Seitz
Richard Szeliski
Brian Curless
Date:
In presenting this dissertation in partial fulllment of the requirements for the doctoral degree at
the University of Washington, I agree that the Library shall make its copies freely available for
inspection. I further agree that extensive copying of this dissertation is allowable only for scholarly
purposes, consistent with fair use as prescribed in the U.S. Copyright Law. Requests for copying
or reproduction of this dissertation may be referred to Proquest Information and Learning, 300
North Zeeb Road, Ann Arbor, MI 48106-1346, 1-800-521-0600, to whom the author has granted
the right to reproduce and sell (a) copies of the manuscript in microform and/or (b) printed copies
of the manuscript made from microform.
Signature
Date
University of Washington
Abstract
Scene Reconstruction and Visualization from Internet Photo Collections
Keith N. Snavely
Chair of the Supervisory Committee:
Professor Steven M. Seitz
Computer Science & Engineering
The Internet is becoming an unprecedented source of visual information, with billions of images
instantly accessible through image search engines such as Google Images and Flickr. These include
thousands of photographs of virtually every famous place, taken from a multitude of viewpoints, at
many different times of day, and under a variety of weather conditions. This thesis addresses the
problem of leveraging such photos to create new 3D interfaces for virtually exploring our world.
One key challenge is that recreating 3D scenes from photo collections requires knowing where
each photo was taken. This thesis introduces new computer vision techniques that robustly recover
such information from photo collections without requiring GPS or other instrumentation. These
methods are the rst to be demonstrated on Internet imagery, and show that 3D reconstruction
techniques can be successfully applied to this rich, largely untapped resource. For this problem
scale is a particular concern, as Internet collections can be extremely large. I introduce an efcient
reconstruction algorithm that selects a small skeletal set of images as a preprocess. This approach
can reduce reconstruction time by an order of magnitude with little or no loss in completeness or
accuracy.
A second challenge is to build interfaces that take these reconstructions and provide effective
scene visualizations. Towards this end, I describe two new 3D user interfaces. Photo Tourism is a
3D photo browser with new geometric controls for moving between photos. These include zooming
in to nd details, zooming out for more context, and selecting an image region to nd photos of an
object. The second interface, Pathnder, takes advantage of the fact that people tend to take photos
of interesting views and along interesting paths. Pathnder creates navigation controls tailored to
each location by analyzing the distribution of photos to discover such characteristic views and paths.
These controls make it easy to nd and explore the important parts of each scene.
Together these techniques enable the automatic creation of 3D experiences for famous sites. A
user simply enters relevant keywords and the system automatically downloads images, reconstructs
the site, derives navigation controls, and provides an immersive interface.
TABLE OF CONTENTS
Page
List of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii
List of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi
Chapter 1: Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
Chapter 2: Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.1 Reconstructing 3D geometry from multiple images . . . . . . . . . . . . . . . . . 10
2.2 Photo browsing and visualization . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3 Image-based rendering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
Part I: Reconstruction of unordered photo collections . . . . . . . . . . . . . . . . 23
Chapter 3: A scene reconstruction pipeline . . . . . . . . . . . . . . . . . . . . . . . . 25
3.1 Finding correspondence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2 Structure from motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.3 Algorithmic complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.4 Geo-registration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.5 Processing the recovered scene . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.7 Failure modes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
Chapter 4: Skeletal graphs for efcient structure from motion . . . . . . . . . . . . . . 68
4.1 Approximate uncertainty analysis using the image graph . . . . . . . . . . . . . . 71
4.2 Building G
I
and G
P
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.3 Computing the skeletal set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.4 Results and discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
4.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
i
Part II: Navigation and visualization of 3D photo collections . . . . . . . . . . . . 97
Chapter 5: Photo Tourism: an immersive 3D photo browser . . . . . . . . . . . . . . . 99
5.1 Rendering scenes and transitions in Photo Tourism . . . . . . . . . . . . . . . . . 101
5.2 Navigation in Photo Tourism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.3 Augmentation and enhancement of scenes . . . . . . . . . . . . . . . . . . . . . . 113
5.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
5.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
Chapter 6: Finding paths through the worlds photos . . . . . . . . . . . . . . . . . . 120
6.1 Related work on 3D navigation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
6.2 System overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
6.3 Viewpoint scoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
6.4 Scene-specic navigation controls . . . . . . . . . . . . . . . . . . . . . . . . . . 131
6.5 Scene viewer and renderer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
6.6 Path planning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
6.7 Appearance stabilization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
6.8 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
6.9 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
Chapter 7: Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
7.1 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
Appendix A: Estimating the focal length of a digital photo from Exif tags . . . . . . . . 182
Appendix B: Complexity analysis of incremental structure from motion . . . . . . . . . 184
Appendix C: Line segment reconstruction . . . . . . . . . . . . . . . . . . . . . . . . . 187
Appendix D: Photo credits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
ii
LIST OF FIGURES
Figure Number Page
1.1 The rst two pages of results of a Flickr search for Trafalgar Square. . . . . . . . 2
1.2 Screenshot from Googles Street View. . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 3D reconstructions from Internet photo collections. . . . . . . . . . . . . . . . . . 6
1.4 Photo Tourism. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.5 Finding paths through a photo collection. . . . . . . . . . . . . . . . . . . . . . . 8
2.1 Traditional image-based modeling results. . . . . . . . . . . . . . . . . . . . . . . 11
2.2 Screenshots from WWMX and Flickr Maps. . . . . . . . . . . . . . . . . . . . . . 18
2.3 Comparison of a computer-generated rendering with a photograph. . . . . . . . . . 19
2.4 Screenshots from the Aspen Moviemap . . . . . . . . . . . . . . . . . . . . . . . 21
3.1 Block diagram of the structure from motion pipeline. . . . . . . . . . . . . . . . . 28
3.2 Difference of Gaussians image lter. . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.3 Example set of detected SIFT features. . . . . . . . . . . . . . . . . . . . . . . . . 30
3.4 Image connectivity graph for the Trevi Fountain. . . . . . . . . . . . . . . . . . . 34
3.5 Illustration of reprojection error. . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.6 Incremental structure from motion. . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.7 Accuracy of focal lengths obtained from Exif tags. . . . . . . . . . . . . . . . . . 42
3.8 Sparsity pattern of the Hessian matrix for a structure motion problem. . . . . . . . 45
3.9 Example registration of cameras to an overhead map. . . . . . . . . . . . . . . . . 50
3.10 Sample reconstructed Internet photo collections. . . . . . . . . . . . . . . . . . . . 57
3.11 More sample reconstructed Internet photo collections. . . . . . . . . . . . . . . . . 58
3.12 Sample reconstructed personal photo collections. . . . . . . . . . . . . . . . . . . 59
3.13 Three images from the nskulla dataset. . . . . . . . . . . . . . . . . . . . . . . . . 62
3.14 Incorrect interpretation of the Florence Duomo. . . . . . . . . . . . . . . . . . . . 62
3.15 Erroneous matches between images of the Arc de Triomphe. . . . . . . . . . . . . 63
3.16 Example of Necker reversal. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.17 Cascading errors in the Colosseum. . . . . . . . . . . . . . . . . . . . . . . . . . . 65
4.1 Image graph and skeletal graph for the Stonehenge data set. . . . . . . . . . . . . . 69
iii
4.2 Dominating sets and t-spanners . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.3 Modeling covariance with an image graph . . . . . . . . . . . . . . . . . . . . . . 75
4.4 Pair graph construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.5 Edge importance. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.6 Reconstruction of Stonehenge using the skeletal set algorithm. . . . . . . . . . . . 88
4.7 Reconstruction of the interior of St. Peters using the skeletal set algorithm. . . . . 88
4.8 Reconstruction of the Pantheon using the skeletal set algorithm. . . . . . . . . . . 90
4.9 Several views of the Pisa1 reconstruction. . . . . . . . . . . . . . . . . . . . . . . 91
4.10 Overhead photo and view of the Pisa2 reconstruction. . . . . . . . . . . . . . . . . 91
4.11 Views of the Trafalgar Square reconstruction. . . . . . . . . . . . . . . . . . . . . 92
4.12 Views of the Statue of Liberty reconstruction. . . . . . . . . . . . . . . . . . . . . 93
4.13 Stretch factor analysis for the St. Peters data set. . . . . . . . . . . . . . . . . . . 94
4.14 Stretch factor analysis for the Temple data set. . . . . . . . . . . . . . . . . . . . . 95
5.1 Screenshots from the Photo Tourism viewer. . . . . . . . . . . . . . . . . . . . . . 101
5.2 Transitions between images. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.3 Frames from an example image transition. . . . . . . . . . . . . . . . . . . . . . . 106
5.4 Delaunay triangulation of an image. . . . . . . . . . . . . . . . . . . . . . . . . . 106
5.5 Zooming controls. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
5.6 Object-based navigation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.7 A stabilized slideshow. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
5.8 A registered historical photo. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
5.9 Example of annotation transfer. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
6.1 Paths through the Statue of Liberty photo collection. . . . . . . . . . . . . . . . . 122
6.2 Illustration of angular deviation. . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
6.3 Scoring a candidate orbit. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
6.4 An orbit dened by an orbit axis and a camera. . . . . . . . . . . . . . . . . . . . 135
6.5 Density function and detected orbits for the Pantheon data set. . . . . . . . . . . . 137
6.6 Panoramas detected in the Pantheon. . . . . . . . . . . . . . . . . . . . . . . . . . 138
6.7 Canonical views detected in the Pantheon data set. . . . . . . . . . . . . . . . . . . 139
6.8 Screenshot of the Pathnder user interface. . . . . . . . . . . . . . . . . . . . . . . 141
6.9 Orbiting a home-made object movie of Mark Twain. . . . . . . . . . . . . . . . . . 142
6.10 Detected orbits in the Statue of Liberty and Venus data sets. . . . . . . . . . . . . . 143
6.11 Proxy planes should intersect the orbit point. . . . . . . . . . . . . . . . . . . . . . 144
6.12 Orientation of proxy planes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
iv
6.13 Blending of images during compositing. . . . . . . . . . . . . . . . . . . . . . . . 146
6.14 Using path planning to compute a camera transition. . . . . . . . . . . . . . . . . . 149
6.15 Computing image similarity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
6.16 Similarity mode and color compensation. . . . . . . . . . . . . . . . . . . . . . . 155
6.17 Switching between day and night in the Trevi Fountain. . . . . . . . . . . . . . . . 156
6.18 Personal photo tour of the Pantheon. . . . . . . . . . . . . . . . . . . . . . . . . . 158
6.19 Free-form paths around the Pisa Duomo and Stonehenge. . . . . . . . . . . . . . . 159
7.1 Image connectivity graph for 20,000 images of Rome. . . . . . . . . . . . . . . . . 165
v
LIST OF TABLES
Table Number Page
3.1 Data sets and running times for the structure from motion pipeline. . . . . . . . . . 55
3.2 Running times for each stage of the structure from motion pipeline. . . . . . . . . 55
4.1 Data sets and running times for the skeletal set algorithm. . . . . . . . . . . . . . . 89
vi
ACKNOWLEDGMENTS
I wish to thank Steve Seitz and Rick Szeliski for their guidance, inspiration, and support during
my graduate studies. I have learned a great deal from them about research, computer vision, and
otherwise, and I deeply cherish the time I have spent working with them. I have also beneted
greatly from knowing and working with Brian Curless, Sing Bing Kang, Michael Cohen, and Larry
Zitnick while at the University of Washington, and with Greg Andrews and Saumya Debray while
an undergraduate at the University of Arizona.
I would also like to thank the numerous current and former UW students and postdoctoral fel-
lows who have helped make my time at the UW such a pleasure. Thanks especially to my collabora-
tors Li Zhang, Pravin Bhat, Ian Simon, Michael Goesele, Colin Zheng, Rahul Garg, Ryan Kaminsky,
and Sameer Agarwal, and to Kevin Chiu and Andy Hou for their help with the Pathnder project.
Very special thanks as well to fellow UW students Craig Prince, Alan Liu, and Tyler Robison for
their camaraderie, and for helping me to move on so many occasions.
Finally, I wish to thank my parents for their love and support.
vii
DEDICATION
to Yihui
viii
1
Chapter 1
INTRODUCTION
A photograph is a window into the world. A good photo can depict a place or an event in
vivid richness and detail, and can give the viewer a taste of being there. However, a single photo
has a xed boundary in space and time, and, unlike with a real window, we cannot simply adjust
our perspective to see what is beyond the frame. Part of the art of photography is working within
these limitations to create works of beauty, by capturing just the right moment with just the right
composition. But photographs are taken for all sorts of reasons in addition to purely artistic ones.
Photos can document and communicate information about people, places, and things in our world,
and are widely used in commercial advertising, classied advertisements, and tourism. At a more
personal level, photos are used to capture moments in our lives, so that we can later relive memories
and share stories with others. For these kinds of applications, where the aim is to capture and convey
information, the limitations of a photograph are more signicant. For instance, it can be difcult or
impossible to give a complete understanding of a large, complex spacea house, Trafalgar Square,
the Louvre, the Grand Canyon, the entire city of Romeor to document your own visit to one of
these places, with a single photo.
Of course, there is no reason to limit oneself to a single photo. This may have been the case in
the early days of photography, when taking a photo involved unwieldy equipment and a painstaking
development process, but the invention of lm, pioneered by George Eastman in the last decades of
the 19th century, led to cheap, easy-to-use cameras and the advent of the snapshot. The past fteen
years have seen another revolution in photography, as digital cameras have steadily replaced lm,
making it easy and inexpensive for a single person to take thousands of photos at a time. The digital
age has also radically changed the way in which we store, view, and share photos. Where we once
stuffed shoeboxes and photo albums with hundreds of photos, we can now store tens of thousands of
photos on a hard drive, and organize themon a computer with tools like Picasa [111] and iPhoto [71].
During the same time, the growth of the Internet has taken this trend even further. We can now store
2
Figure 1.1: The rst two pages of results of a Flickr search for Trafalgar Square.
virtually unlimited numbers of photos online on sites such as Flickr [47], SmugMug [135], Picasa
Web Albums [112], and Facebook [43]. Not only do these sites give us access to our photos from
anywhere, but they also give us access to everyones photos, compounding the number of photos we
have at our ngertips many times over.
Thus, if we are selling our house and decide to advertise it online, we need not use just a single
photowhy not take a hundred? Similarly, if we want to understand what it is like to experience
a famous place like Trafalgar Square, we can go to Flickr, search for Trafalgar Square, and nd
tens of thousands of photos captured from every imaginable viewpoint, by thousands of different
people, under many different weather conditions, at many times of day and night, and during all
times of the year. We can also nd photos of the Square during different events: protests (search
for Trafalgar Square protest), a visit by the Sultans Elephant (Trafalgar Square elephant), or
Christmas celebrations (Trafalgar Square Christmas). We can nd similarly rich collections of
photos for almost every famous world site. The Internet, with its billions and billions of photos,
3
represents an unprecedented visual record of our world.
How can we use these vast, rich photo collections to effectively communicate the experience
of being at a placeto give someone the ability to virtually move around and explore a famous
landmark, to see what it looks like at sunrise and sunset or at night, to revisit different events; in
short, to convey a real understanding of the scene? Unfortunately, the availability of vast image
collections and the ability to capture them are alone not enough to create these kinds of experiences.
In fact, with current tools, the more photos there are, the harder it often is to make sense of them.
Most photo browsing tools treat photos as independent views of events or scenes, perhaps grouped
together or labeled in meaningful ways, but otherwise visually disconnected. Such tools do not
attempt to exploit or convey the rich common structure that exists among photos. For instance, the
Trafalgar Square photos from Flickr shown in Figure 1.1 are arranged in a grid of thumbnails, the
default way of presenting search results in Flickr, and a nearly ubiquitous display idiom in photo
browsing software. This kind of presentation does not facilitate the understanding of the space as a
whole. Nor does it really even scale well as a way of simply displaying photos. Only about twenty-
ve thumbnails are shown on a page, so to see all the photos we would have to view hundreds of
pages of results. Finally, the fact that these photos are submitted by many different people is a source
of richness and variety, but also a source of disorder. The photos are largely unstructured, and not
presented in a particularly meaningful order. Certain tasks, such as nding a specic viewpoint or a
detail of a particular object, can be quite difcult.
In contrast, several commercial software applications have started to present large photo collec-
tions of urban cityscapes organized in a much more structured way, making it much easier to explore
the underlying scene [58, 159, 41]. For instance, Google Street View simulates the experience of
walking down the streets of major cities by displaying omnidirectional photos taken at intervals
along every city street, as shown in Figure 1.2. Such web applications, combining the high visual
delity of photos with simplied 3D navigation controls, are useful for recreating the experience of,
say, walking around the Haight Ashbury district of San Francisco. However, capturing these expe-
riences typically requires special camera hardware, careful attention to the photo capture process,
and time-consuming post-processing and quality control [154], hence they are not easy for casual
photographers to create (in fact, Everyscape, a company with a related application, recommends
special training courses and equipment to people who want to create their own content [42]). Part
4
Figure 1.2: Screenshot from Googles Street View, showing an intersection in downtown San Fran-
cisco. By clicking on the arrows, a user can virtually move through the city streets.
of the difculty arises from the need for accurate location information for every photo to put each
in its proper context.
Is it possible to apply these kinds of 3D scene visualization tools to personal or Internet photo
collections? If so, we could not only create 3D experiences for all the worlds sites from existing im-
agery, but we could also potentially create much more powerful tools for browsing photos of places.
Unfortunately, the vast majority of these casually acquired collections lack location information, a
prerequisite for assembling the photos into a coherent whole.
My solution to this problem is to turn to automatic computer vision techniques. Researchers
in the eld of multi-view geometry have made great progress towards automatic recovery of ne-
grained 3D location information for photo collections through the analysis of the photos them-
selves [68]. However, most previous systems have focused on reconstructing controlled image col-
lections: sequences that are well-ordered in time (such as video) and which consist of images taken
under similar conditions and with the same camera. Applying computer vision techniques to the
large, fundamentally unordered collections typical of the Internettaken by many different people,
under diverse conditions, and with myriad camerasraises many new challenges.
5
Furthermore, adding location information to photo collections is by itself insufcient for scene
visualization: we also need intuitive, interactive interfaces for exploring these scenes. There are
several challenges faced in the design of such interfaces. First, unlike with Google Street View,
where photos are taken at regular intervals, personal or Internet collections are typically an un-
structured soup of photos. Nevertheless, the navigation controls should still be intuitive and exhibit
regularity. Second, such controls should make it easy to nd and explore the interesting parts of a
scene, particularly for tourist sites.
1.1 Contributions
There are two fundamental research challenges identied above: (1) the computer vision challenge
of reconstructing 3D information from Internet photo collections, and (2) the computer graphics
challenge of using this 3D information to enable new ways to visualize and explore famous sites.
In this thesis, I present a range of new algorithms and techniques in both vision and graphics that
address each of these challenges.
Reconstruction of Internet photo collections. To address the computer vision challenge, I have
developed new 3D reconstruction algorithms that operate on large, diverse image collections. These
algorithms recover both camera pose and scene geometry and demonstrate, for the rst time, that
3D geometry can be reliably recovered from photos downloaded from the Internet using keyword
search. Hence, they enable reconstruction of the worlds famous sites from existing imagery on the
web. A few example reconstructions are shown in Figure 1.3, and many more appear in Chapters 3
and 4.
Two major issues faced in handling Internet photo collections are robustness and scale. We
desire computer vision techniques that work reliably across a wide variety of scenes and photo
collections, and on collections with large numbers of photos. Robustness is a concern because
there are often ambiguities and degeneracies that crop up in multi-view geometry, and because
there is no perfect algorithm for nding correspondences between images. These problems are
heightened when handling unstructured collections, which lack prior information on the structure
of the image sequence. Chapter 3 describes a robust reconstruction pipeline that recovers scene and
camera geometry incrementally, carefully building a reconstruction a few images at a time. The
6
Figure 1.3: 3D reconstructions from Internet photo collections. My computer vision system takes
large collections of photos from the Internet (sample images shown at top) and automatically recon-
structs 3D geometry (bottom). The geometry includes camera informationwhere each photo was
taken and which direction it was lookingas well as a point cloud of the scene. In these images of
reconstructions the recovered cameras are shown as black wireframe pyramids, and the scene is ren-
dered as a point cloud. From left to right: the Statue of Liberty (reconstructed from 388 photos from
Flickr), Half Dome in Yosemite National Park (reconstructed from 678 photos), and the Colosseum
in Rome (reconstructed from 1,135 photos).
pipeline starts with a pair of images that exhibit strong geometry, giving the reconstruction a solid
foundation, then conservatively adds new information until the entire scene model is reconstructed. I
demonstrate this pipeline on a number of different photo collections, including urban environments,
mountains, and indoor scenes. This system has been used to reconstruct well over one hundred
photo collections, by myself and other researchers. In addition, Photosynth [109], a commercial
photo reconstruction system released by Microsoft and based largely on the work described here,
has been applied to thousands of different photo collections by many users.
Chapter 4 addresses the issue of scalability by introducing skeletal sets, an approach for im-
proving the efciency of reconstruction by reducing the amount of redundant information. This
technique represents the information present in a photo collection as a graph, and computes a much
sparser skeletal subgraph that preserves the overall structure of the graph and bounds the amount
7
Figure 1.4: Photo Tourism. A screenshot from the Photo Tourism interface, showing a reconstruc-
tion of the Trevi Fountain. Information about the currently displayed photo, as well as geometric
search tools for nding related photos, appear in panes on the left and bottom of the screen.
of information lost. The hard computational effort of scene reconstruction is then applied only to
the skeletal subgraph; the remaining images can be added in later using simple pose estimation
techniques. The skeletal sets algorithm reconstructs scenes using many fewer images and reduces
reconstruction time by an order of magnitude for large image collections, with little or no loss in
reconstruction quality.
Visualization of 3D photo collections and scenes. This thesis also describes computer graphics
and interaction techniques for taking these reconstructed scenes and creating new ways to browse
photo collections and to visualize the world. For this problem, two primary concerns are (1) how
to provide the user with effective navigation controls for exploring scenes and (2) how to display or
render scenes using the imagery available in large photo collections. I have designed two interfaces
that address these challenges in different ways.
Chapter 5 describes Photo Tourism, an immersive 3D photo browser. Photo Tourism situates the
user in the 3D scene among the photo collection, as shown in Figure 1.4. The interface provides
new geometric controls for moving through the collection: users can zoom in and nd details of
a given photo, zoom out to see more context, select a region of an image to nd a good photo of
a specic object, and move in different directions (e.g., to the left or right) to view more images.
8
(a) (b) (c)
Figure 1.5: Finding paths through a photo collection. (a) Several Flickr images of the Statue of
Liberty from a collection of 388 input photos. (b) Reconstructed camera viewpoints for this collec-
tion, revealing two clear orbits, shown superimposed on a satellite view. (c) A screenshot from the
Pathnder user interface. The two orbits are shown to the user as thumbnails at the bottom of the
screen. Clicking on a thumbnail moves the user on an automatic path to the selected control.
The interface also provides powerful annotation features. A user can label an object in one photo
and have that label propagate to all other images in the collection. On the rendering side, I intro-
duce new techniques for creating attractive 3D transitions between photos and generating stabilized
slideshows that make it easier to see how a scene changes over time.
Chapter 6 describes the second system, Pathnder. While Photo Tourism considers the prob-
lem of photo browsing, Pathnder takes advantage of large Internet photo collections to create an
enhanced interface for exploring the scenes themselves. These photo collections, captured by many
people, are an extremely valuable source of information for determining good controls for navigat-
ing through a scene, as they represent samples of how people actually experienced the scene, where
they stood, and what views they found interesting. For instance, Figure 1.5(b) shows a reconstructed
collection of Flickr photos of the Statue of Liberty. Even if we knew nothing about the structure or
content of this scene, the photos are a powerful indicator of the presence of an interesting object,
as they are almost all trained on a single focal point (the center of the statue), and are organized
into orbits around that point. These orbits are natural controls for exploring the scene. Pathnder
analyzes the distribution of reconstructed photos to derive such controls, and presents them to the
user in a uid, game-like navigation interface, shown in Figure 1.5(c). Pathnder also introduces a
new rendering technique that continuously selects and warps input photos as the user moves through
a scene.
9
Taken as a whole system, my work allows a user to simply specify a search term (Trevi Foun-
tain, Notre Dame Cathedral) and get back an immersive 3D experience of that place. The system
automatically downloads related photos from sites like Flickr, creates a 3D reconstruction, computes
a set of custom navigation controls, and provides these controls to the user in an interactive scene
browser.
This thesis is divided into two main parts. Part I describes the new computer vision algorithms
required to reconstruct 3D geometry from large Internet photo collections. Part II describes the
computer graphics, 3D navigation, and user interface innovations that enable new ways to explore
photo collections and scenes in 3D. Chapter 7 discusses several directions for future work. Before
describing the new algorithms and techniques, I rst review related work in computer vision and
computer graphics in Chapter 2.
10
Chapter 2
RELATED WORK
The work described in this thesis ultimately has two intertwined goals: (1) to make it easier
to browse and explore large collections of photos related by place, and (2) to create new ways
of virtually exploring real-world scenes through image collections, a version of the image-based
rendering problem. Both of these goals rely on the ability to reconstruct geometry automatically
from large, unstructured image collections. In this chapter, I review related work on each of these
problems: 3D reconstruction from multiple images (Section 2.1), photo browsing (Section 2.2), and
image-based rendering (Section 2.3).
2.1 Reconstructing 3D geometry from multiple images
Multi-view geometry, or the recovery of 3D information from multiple images of a scene, has long
been an active research area in computer vision and graphics, and has inspired a wide variety of
different approaches and algorithms. In computer graphics, the aim is often to acquire geome-
try suitable for high-quality rendering in a traditional graphics pipeline, i.e. closed, polygonal 3D
models, such as those shown in Figure 2.1, with detailed appearance information (texture maps,
bi-directional reectance distribution functions, etc.). This problem is known as image-based mod-
eling. Image-based modeling techniques tend to be manual or semi-automatic, as graphics appli-
cations often demand a level of photorealism that requires ne control over the quality of the 3D
models.
In contrast, I show that sparse scene geometry, recovered completely automatically from a photo
collection, is often sufcient to produce attractive and comprehensible scene renderings and tran-
sitions between photographs. In my approach, the key problem is to robustly and efciently
and completely automaticallyrecover camera and sparse scene geometry from large, unorganized
photo collections.
11
(a) (b) (c)
Figure 2.1: Traditional image-based modeling. Images from Facade, a semi-automatic image-based
modeling system [32]. (a) Images are annotated with lines indicating geometric structures, and
(b) these annotations are used to recover a polygonal 3D model. (c) The model (and associated
appearance information) can then be used to render new views of the scene.
2.1.1 Structure from motion
The precursor to modern multi-view geometry techniques is the eld of photogrammetrythe sci-
ence of taking 3D measurements of the world through photographswhich began to develop soon
after the advent of photography.
1
The key problem in photogrammetry is to determine the 3D lo-
cation of a point in a scene from multiple photographs taken from different vantage points. This
problem can be solved using triangulation if the pose (the position and orientation in the world) of
each photo is known. Conversely, if we have a set of points with known 3D coordinates, the pose
of photos that observe those points can be determined using a process known as resectioning
2
or
camera pose estimation. But what if we know neither the camera poses nor the point coordinates?
Estimating both at once seems like a chicken-and-egg problem. The traditional photogrammetric
1
Although an important technique in photogrammetry, triangulation, has been known since antiquity.
2
The word resection is by analogy with computing the intersection of rays. To triangulate a 3D point, we can
observe it from two or more known locations, and intersect the appropriate rays. Conversely, another way to nd the
coordinates of a 3D point is to stand at that point and observe two or more points with known coordinates. We then
resect the rays to the known points to nd our location. This is a simplied form of camera pose estimation.
12
solution to this problem is to place a set of reference ducial markers at known 3D positions in
the scene, manually identify them in each image, and use these to recover the poses of the cam-
eras through resectioning. Once the camera poses are known, additional unknown points can be
identied in each image and triangulated.
Over time, the requirements of this traditional approach have steadily been relaxed, due to ad-
vances in both computer vision and photogrammetry. Correspondences between image points can
often be computed completely automatically without the need for ducials, and algorithms have
been developed for computing camera pose and scene structure simultaneously, without requiring
either to be known a priori. This is known in computer vision as the structure from motion (SfM)
problem.
The simplest version of the SfM problem, involving just two images, has been extensively stud-
ied. Nearly one hundred years ago, Kruppa proved that for two views with ve point correspon-
dences, the camera poses and 3D point locations can be determined (up to a similarity transform)
[81], and based on this result several ve-point algorithms for estimating two-view geometry have
recently been developed [102, 86]. The mathematical and algorithmic aspects of the three-view
problem have also received a great deal of attention [68]. For larger numbers of images, however,
the problem becomes more difcult. For some specic scenarios the multi-image SfM problem can
be solved exactly, but for the general case no such closed-form solutions exist, and a wide variety of
different algorithms have been proposed.
One common multi-image scenario is that the images are given in a particular sequence, such
as the frames of a video or images captured by a robot moving through a building. Assuming
that one can match corresponding points between consecutive images, a simple way to reconstruct
such a sequence is to use the ve-point algorithm to recover the poses of the rst two frames and
triangulate their common points, then resection the next frame, triangulate any new points, and
repeat: a form of visual dead reckoning. With well-calibrated cameras, this technique can perform
well, particularly on short image sequences [102]. However, the main problem with this approach is
that errors in pose estimates will accumulate over time. Better results can be obtained by propagating
information forwards and backwards within a window of frames, e.g., through Kalman ltering and
smoothing [94].
These local techniques are very efcient, and work well for camera trajectories that are simple,
13
i.e., that do not double-back or form loops. For complex paths, much more accurate geometry can be
obtained by using global optimization to solve for all the camera poses and point positions at once.
In certain specic cases, this can be done in closed form using factorization. Originally proposed by
Tomasi and Kanade [148], factorization methods are based on the observation that for orthographic
cameras, the 2nmobservation matrix W, formed by stacking rows of zero-mean observations for
each camera, can be factored into the product of a 2n 3 matrix R whose rows represent camera
rotations, and a 3 m matrix S whose columns are the 3D positions of the points. The factoriza-
tion W = RS can be computed using singular-value decomposition. Factorization methods have
been extended to more general camera models, such as paraperspective [114] and perspective cam-
eras [141, 23], but have a number of drawbacks. First, for perspective camera models, factorization
methods minimize an algebraic error function with no direct geometric interpretation. Second, in or-
der to directly factor the observation matrix W, all of its entries must be known, i.e., all points must
be visible in all images. In most real-world situations, someindeed, mostof the observations
are unknown, due to occlusions or missing correspondences. When there is missing data, no known
closed-form solution to factorization exists, although several iterative schemes have been proposed
[148, 67]. Finally, it is difcult to incorporate priors and robust error functions into factorization
methods [16], extensions which can be critical for handling noisy correspondences contaminated
with outliers.
Bundle adjustment [150] is an alternative to factorization that can easily handle missing data
and accommodate priors and robust objective functions. Bundle adjustment
3
seeks to minimize a
geometric cost function (e.g., the sum of squared reprojection errors) by jointly optimizing both the
camera and point parameters using non-linear least squares [143]. Given a known measurement
noise distribution, bundle adjustment is the statistically correct way to nd the parameters, and is
regarded as the gold standard for performing optimal 3D reconstruction [68]. However, bundle
adjustment has no direct solution and involves minimizing a non-linear cost function with poten-
tially many local minima. Therefore, bundle adjustment requires careful initialization, which can be
difcult to obtain.
Another disadvantage to bundle adjustment is that it can be quite slow for scenes with large num-
3
The name bundle adjustment stems from the idea of adjusting bundles of light rays from scene points to camera
centers to align them with the observations.
14
bers of cameras and points. While bundle adjustment problems have a particular form that makes
them amenable to sparse matrix techniques [143], even approaches that take advantage of sparsity
(e.g., [15] and [89]) can become very slow when the number of cameras is large. Triggs et al.found
empirically that sparse bundle adjustment methods can have approximately cubic complexity in the
number of cameras [150].
Faster SfM methods based on bundle adjustment have been designed for ordered image se-
quences, such as video, where several simplifying assumptions can be made. For instance, a com-
mon assumption is that the information gained by adding a new video frame only affects a small
number of frames immediately preceding the new frame. As a result, only a constant number of
camera parameters, and the points they observe, must be adjusted whenever a new frame is added.
This assumption is the basis for several real-time systems for SfM from video [116, 103, 25, 115].
Engels et al. [39] analyzed the inherent tradeoff between speed and accuracy for this type of ap-
proach when considering how many frames to optimize, and Steedly and Essa proposed a more
principled algorithm that considers the ow of new information in a system when determining the
set of parameters to update [140].
Other bottom-up systems improve performance using a divide-and-conquer approach. In the
work of Fitzgibbon and Zisserman [46], a video is divided into small, three frame subsequences,
which are reconstructed independently, then merged into progressively larger reconstructions. Nist er
[101] generalized this approach to partition the video into contiguous subsequences of irregular
length, resulting in better-conditioned reconstructions. Steedly et al. [139] proposed a technique
for rening an initial reconstruction by segmenting it into clusters using spectral partitioning, then
treating each cluster as a single rigid body during optimization, dramatically decreasing the number
of parameters to estimate. Ni et al., also use segmentation to enable a fast, out-of-core bundle
adjustment technique [100].
These bottom-up techniques are effective because of the spatial continuity of video; dividing
a video into contiguous subsequences is a reasonable way to partition the frames. Unfortunately,
photo collections, especially those found through Internet search, often have much less structure
than video sequences. Typically, search results are presented in order of relevance, or other criteria,
such as date, which have little to do with spatial continuity. It can therefore be difcult to apply fast
SfM methods designed for video to these kinds of collections. Recently, a few researchers, including
15
Schaffalizky and Zisserman [124], Brown and Lowe [15], and Vergauwen and Van Gool [153] have
considered the problem of applying SfM to such unordered photo collections. Martinec et al. [91],
presented a way to use estimates of relative pose between pairs of cameras to robustly estimate a
set of globally consistent camera poses. However, these techniques have largely been demonstrated
on relatively small and controlled sequences of photos captured by a single person. Internet photo
collections can be orders of magnitude larger than collections that have been considered previously,
and have fundamentally different properties due to the distributed nature of their capture. In my
work, I demonstrate new algorithms that can robustly reconstruct Internet photo collections of up to
several thousand images.
2.2 Photo browsing and visualization
Over the past few years, many researchers have sought better computational tools for organizing,
visualizing, and searching through large collections of digital photos. Most have focused on large
personal photo collections, but others have considered the problem of Internet collections. In this
section, I survey previous work in the area of photo browsing and visualization.
Large-scale collections of media can be very challenging to interact with and to visualize. A
person can easy accumulate thousands of photos, far more than can be viewed in a comprehensible
way all at once. As the volume of media increases, tasks such as getting an overview of a collection,
organizing or structuring a collection, nding a specic photo, or nding a set of photos with certain
characteristics become more difcult and require more sophisticated tools. Many proposed tools in
the photo browsing literature can be decomposed into two components: (1) a method to lter or
sort the images and (2) a method to display and interact with the results of the sorting and ltering
operations. Many of these sorting and display techniques have found their way into commercial
photo browsers, such as Googles Picasa [111] and Apples iPhoto [71].
Sorting and ltering photos. Many approaches to sorting and ltering focus on either (a) making
it easy to organize the photos into meaningful structures or (b) creating better tools for photo search.
For many people, the default method for organizing photos is to use the le system to arrange the
photos in a directory tree, where photos in each node can be sorted by attributes such as time. Some
tools, such as the PhotoTOC system [113], attempt to automatically create a meaningful hierarchy
16
by clustering photos into events based on time stamps. Girgensohn et al. [53] also cluster photos
into events, but provide multiple, parallel hierarchies based on the people and places featured in the
images.
Other work addresses the problem of organization through search. The PhotoFinder system [75]
represents the photo set as a database and presents an intuitive interface for formulating complex
database queries (involving attributes such as time, people, location, and rating). Tagging, or anno-
tating photos with text, has also emerged as a useful tool for managing personal photo collections
[82]. Tags are a exible way to organize photos in a way that makes sense to a particular user,
without enforcing a strict hierarchical structure; tags are also easily searchable. Several researchers
have sought easy ways to apply tags to large numbers of personal photos, for instance through
drag-and-drop interfaces [129], automatic tagging based on contextual information [29], or efcient
selection mechanisms [37]. Others have addressed the problem of tagging Internet photos, through
collaborative tools such as LabelMe [123] or games such as the ESP Game [155]. Tagging is also
an important capability of many Internet photo-sharing sites, including Flickr [47] and Picasa Web
Albums [112].
Displaying photos. Other research has focused on the display of large photo collections. Most
photo browsing software displays photos using grids of thumbnails or slideshows, which do not
always scale well. Several other 2D display techniques have been proposed, such as calendar and
scatter plot views [75], alternative thumbnail layouts [9], and zoomable browsers including Pho-
toMesa [9] and Seadragon [127]. The MediaBrowser [37] also has a timeline mode where photos
are displayed in 3D stacks.
Using image content. Another line of research has looked at using the content of photos to im-
prove the search and display of image collections. While a complete survey of content-based search
tools is outside the scope of this thesis, one particularly relevant project is the Video Google system
[133], which enables a user to nd all the frames of a video in which a certain object appears by
selecting it in a single frame.
17
Using location information. More relevant to my research is work that uses location information
to assist in organizing, querying, and displaying photo collections. In most previous work, absolute
location, obtained through GPS or manual tagging, is assumed to be provided in advance for each
photo; such photos are referred to as being geo-referenced or geo-tagged. Some approaches have
used location to aid in clustering photos in a personal collection. For instance, Naaman et al. [98]
recognize that an event (such as a birthday party) has an extent in both space and time, and present
a system, Photocompas, that uses time and location to compute better event clusterings.
In the LOCALE project [97], Naaman et al. explore how location can be used to enable au-
tomatic tagging of photos. LOCALE uses a database of tagged, geo-referenced photos to transfer
tags to new geo-referenced photos based on proximity. Large Internet photo collections have also
been used to infer semantic information about scenes. For instance, several researchers have used
clustering to segment collections of tagged photos into separate places (e.g., Coit Tower and
Transamerica Building), and then infer tags describing each place, using location information
alone [72], visual content [130, 73], or both [77, 118]. These systems also select a set of summary
images useful for getting a quick overview of a collection.
Others have integrated maps into tools for interacting with geo-tagged photos. One such project
is GTWeb [138], which produces a webpage summary of a trip by merging time-stamped photos
with information froma GPS track, a digital map, and a landmark database. In the domain of Internet
photo collections, the World-Wide Media eXchange (WWMX) [149] allows users to share geo-
referenced photos and search for and view photos on a map. Large-scale commercial applications,
such as Flickr [47] and Google Earth [57], have recently adopted similar interfaces for visualizing
geo-referenced photos. Screenshots from WWMX and Flickr Maps are shown in Figure 2.2.
Relatively little work has been done on immersive photo browsing tools, where photos are situ-
ated directly in a 3D environment. This is perhaps partly because in order for such tools to be most
effective, each photo should be localized well enough to be visually aligned with its subject to within
a few pixels (i.e., pixel-accurate camera pose is needed), which requires more positional accuracy
than can be reliably obtained from GPS or manual methods. (In addition, such applications require
accurate orientation information, requiring an electronic compass or similar device.) Nevertheless,
a few projects have taken steps towards immersion. Realityythrough [93] is a method for browsing
a live scene using a set of location-aware video cameras. A user can move around the scene by
18
Figure 2.2: Two 2D interfaces for browsing photos with location. Left: the World-Wide Media
eXchange (WWMX); right: Flickr Maps
selecting different videos, and smooth transitions provide context when moving from one video to
another. While Realityythrough allows a user to explore the scene by selecting among video feeds,
the work of Kadobayashi and Tanaka [74] enables the opposite functionality: a user moves a virtual
camera through a 3D model of a real scene (such as a archaelogical site), then queries an image
database to nd photos taken near the current virtual camera position.
Another tool, PhotoWalker [146], allows a user to create a walkthrough of a scene from a col-
lection of photos by manually linking and specifying morphs between pairs of photos. When a
user is viewing the collection, links are displayed as highlighted regions superimposed on an image;
clicking on one of these regions activates a morph to the next photo. This tool can give the illusion
of a 3D world, although only 2D morphs are used. Similarly, Sivic et al. use 2D morphs between
images unrelated by a specic place (for instance, generic city images) to create the illusion of a
large virtual space [132]. Other projects have made more explicit use of 3D information, including
the 4D Cities project [125], which has the goal of creating an interactive 4D model of the city of
Atlanta from a collection of historical and modern photographs; and Photosynth [109], a web-based
tool for creating and exploring 3D photo collections, which is based on the Photo Tourism project
described in Chapter 5.
These immersive 3D (or pseudo-3D) photo browsers use aspects of 3D navigation, but arguably
do so with the aim of making it easier to browse a photo or media collection. There is also a large
19
Figure 2.3: Comparison of a computer-generated rendering with a photograph. Right: a photograph
of the Paris Las Vegas casino. Left: a screenshot from Microsofts Virtual Earth 3D from roughly
the same viewpoint. While these views show (mostly) the same objects, the photograph is more
vivid and lifelike, capturing the ripples and reections in the water, sharper detail in the buildings,
and realistic shadows and other lighting effects. There are other, more subtle differences as well.
For instance, in the photograph, the color of the sky at the horizon is different from the color closer
to the zenith.
body of work that starts from the opposite direction, where the goal is to recreate virtual 3D versions
of real-world scenes using collections of images or video. This approach to capturing and displaying
the world is known as image-based rendering.
2.3 Image-based rendering
A primary goal of computer graphics is to recreate the experience of being there, giving a user a
sense of being in another place, whether it be a real-world location (such as the Taj Mahal), or a
ctional world (Middle Earth). Two critical aspects of this problem are (a) rendering, or visually
depicting the scene, and (b) navigation, or the mechanisms by which the user moves around in and
explores the scene.
Approaches to rendering and navigation depend fundamentally on the way in which the virtual
scene is created and represented. The traditional computer graphics approach to scene representa-
20
tion, dominant in 3D computer animation and games, is to explicitly model the 3D geometry and
appearance of the scene in full detail. However, creating good 3D models often requires difcult,
time-consuming manual effort, despite the progress in automatic and semi-automatic modeling de-
scribed in Section 2.1. Furthermore, there can be a lack of verisimilitude in renderings produced
using 3D models, especially in real-time applications. Figure 2.3 shows a Virtual Earth [88] ren-
dering of the Paris Las Vegas casino juxtaposed with a photograph taken from the same viewpoint.
While the rendering is quite good, and similar in many respects to the photo, it is still possible to
tell the two apart.
Another very promising approach to creating the experience of being there involves capturing
the appearance of the desired scene through photos or video, then replaying or resampling the cap-
tured images in an interactive viewer. This approach is known as image-based rendering (IBR) and
has inspired a large body of research over the past fteen years. This section reviews prior work in
IBR and related elds.
Moviemaps. A forerunner to the eld of image-based rendering was Andrew Lippmans seminal
Aspen Moviemap project fromthe late 1970s [87]. The Aspen Moviemap used a set of images taken
throughout the town of Aspen, Colorado and stored on laserdisc to create an interactive surrogate
travel application. The goal of the project was, in the words of Lippman, to create so immersive
and realistic a rst visit that newcomers would literally feel at home, or that they had been there
before. [99]
To capture the town on laserdisc, a crew drove a car mounted with cameras through every street
in town, taking photos every ten feet. The cameras were arranged so as to capture views in all
directions. The resulting data was then processed by hand to turn it into an immersive, interactive
travel experience, where a user could virtually drive down the streets of Aspen. The interface was
relatively simple yet intuitive: controls for moving forward and turning were shown on a touchscreen
display (see Figure 2.4). The moviemap was not only an effective new medium in itself but was also
conceived of as a backbone on which a comprehensive audio-visual survey of Aspen could be
spatially organized. [99] Various other media, including still images of every building in town
(shot in both summer and winter), historical photos, documentary video, and sound, were accessible
through hyperlinks from relevant parts of the moviemap.
21
Figure 2.4: Screenshots from the Aspen Moviemap. Left: the interface for virtually driving through
the city. Right: a few still photographs from different seasons hyperlinked to the moviemap.
Although the Aspen Moviemap was an effective means of virtual travel, it required a substantial
amount of time to capture and processmore than a year of effort by a small squadron of at least
a dozen people [99]. Only recently have companies, such as Microsoft and Google, begun to create
such surrogate travel applications on a large scale. Though some advances, such as GPS, have made
the process easier, the preparation, time, and resources required to create such experiences is still
signicant.
The Aspen Moviemap simply plays back existing video frames based on a users actions, re-
sulting in an experience where the user moves discretely between views. Much of the more recent
work in image-based rendering has focused on generating a continuum of new views from a discrete
database of existing images. These techniques have been applied to different scenarios of varying
complexity. The QuickTime VR system [21] renders new views of a scene from a xed position, but
where the camera can rotate and zoom freely; recent work has extended this idea to video panora-
mas [3], multi-viewpoint panoramas [2], and gigapixel images [79]. Approaches based on explicit
sampling of a 4D light eld [84, 59], allow for free motion of a camera within a volume of view-
points. When per-pixel scene geometry or correspondence is known, methods based on warping
and combining images can be used to generate in-between views [20, 95]. An alternative approach,
which also uses scene geometry, is view-dependent texture mapping [32], in which polygons in the
scene are texture mapped with several weighted textures derived from a set of captured images. The
22
weights are based on how close the current viewpoint is to each captured view. Finally, unstructured
lumigraph rendering [17] generalizes many of these previous methods with a framework that can
handle a variety of arrangements of image samples, with or without known geometry.
Researchers have also developed techniques for creating walkthroughs of large, complex spaces.
Arbeeny and Silver [6] created a system for registering video of a scene walkthrough with a 3D
model, allowing for spatial navigation of the video stream. Plenoptic stitching [4] is a method
for generating novel views of a large-scale scene by resampling frames of omnidirectional video
captured on a rough 2D grid throughout the space, allowing a user to walk anywhere within the
captured scene. Other approaches, including Taylors VideoPlus project [147] and Googles Street
View [58], have revisited the moviemaps nearest neighbor approach for resampling images, but
use omnidirectional images to allow for continuous panning and tilting of the camera. Uyttendaele,
et al. [152] propose a similar system using omnidirectional video, and augment their walkthroughs
with additional information, including overhead maps, explicit branching points where the video
stream crosses itself in space, and other media such as video textures.
These recent systems bring us around nearly full circle to the end of Section 2.2, which described
photo browsers that incorporate elements of 3D navigation. Here we have a number of systems for
3D walkthroughs that are constructed from photos or video. While researchers have explored a wide
variety of different design points in image-based rendering (and 3D photo browsing), most IBR sys-
tems are created using carefully captured imagery taken along planned paths through a space. In this
thesis, I focus on large Internet photo collections, which are captured in a highly distributed manner
by uncoordinated individuals, and are thus fundamentally unplanned and unstructured. In Part II of
this thesis, I address the open question of creating effective renderings and navigation controls
and tying these together into a understandable experience of a scenefor such unstructured photo
collections.
23
Part I
RECONSTRUCTION OF UNORDERED PHOTO COLLECTIONS
The ultimate goal of my work is to take a collection of photos related by place, and create an
immersive 3D experience where a user explores the photo collection and the underlying 3D scene.
As a prerequisite to creating this kind of experience, we rst need to recreate a 3D version of
the scene and localize the cameras within that scene. In particular, we need to know where each
photograph was taken and in which direction the camera was pointed, as well as internal camera
settings, such as zoom, which affect how incoming light is projected onto the lm or sensor plane.
I will refer to photos for which such information has been determined as being registered.
How can we register our photos? One solution is to use special hardware. We could equip
our cameras with a Global Positioning System (GPS) device and electronic compass, and use them
to stamp each photo with a position and orientation. Unfortunately, the vast majority of existing
photographs were taken without this kind of specialized hardware, and most digital cameras still
lack such intrumentation.
4
Moreover, the visualization techniques I have developed require pixel-
accurate registration, meaning that if the photograph were retaken from the recorded position and
orientation, the original and new photo should be in alignment within a few pixels. The accuracy of
GPS depends on a number of factors (atmospheric effects, the presence of nearby walls, etc.), but
many manufacturers of GPS receivers report a typical accuracy of around 3-5 meters [1]. Localiza-
tion errors on the order of several meters are much too large to guarantee pixel-accurate registration.
Furthermore, GPS works poorly indoors.
What about internal camera parameters such as zoom? These can be determined by calibrating
a camera, which usually involves taking several photos of a known pattern such as a checkerboard,
as with the calibration systems of Zhang [160] and Bouguet [13]. Again, however, most existing
photographs have been taken with cameras that have not been calibrated, and by people who may
4
A few camera models with built-in GPS, such as the Ricoh 500SE, the Apple iPhone 3G, and the Nikon P6000, are
beginning to appear.
24
not be willing to invest time in calibration.
Because most existing photos lack the necessary metadata for registration, in my work I do not
rely on the camera or any other piece of equipment to provide location information, nor do I assume
that calibration information is available for each image. Instead, I develop a set of algorithms that
derive camera geometry directly from the images themselves using computer vision techniques. In
addition to camera geometry, my system also recovers scene geometry in the form of a sparse set
of 3D points, or point cloud. In this thesis, I refer to the camera and scene geometry recovered
from a photo collection as a reconstruction. Example reconstructions are shown throughout the
next two chapters. The computer vision techniques I use only recover relative camera positions
and scene geometry which are not grounded in an absolute coordinate system (e.g., latitude and
longitude)the system can register all the cameras with respect to each other, but cannot determine
where in the world the reconstruction sits, unlike with GPS.
5
However, this relative pose information
is sufcient for most of the visualization techniques described in Part II. Chapter 3 also describes
several methods for obtaining absolute coordinates.
In the next two chapters, I describe the computer vision algorithms I have developed for re-
constructing camera and scene geometry from unordered photo collections. Chapter 3 gives an
overview of my basic reconstruction pipeline, and presents results on eleven Internet and personal
photo collections. The basic algorithm can operate on collections of up to about a thousand images,
but does not scale well beyond that. To address the scalability problem, Chapter 4 introduces the
idea of the skeletal graph, a concept which can be used to dramatically prune the number of im-
ages required for reconstruction while maintaining guarantees on completeness and accuracy. This
technique allows for much more efcient reconstruction of large photo collections.
5
These reconstructions satisfy a slightly different denition of pixel-accurate than the one described above: recon-
structed 3D points, when projected into a reconstructed camera, lie close to where that 3D point actually appears in the
image.
25
Chapter 3
A SCENE RECONSTRUCTION PIPELINE
My reconstruction pipeline takes an unordered collection of images (acquired, for instance, from
Internet search or a personal photo collection), and produces 3D camera and scene geometry. In
particular, for each input photo, the pipeline determines the location from which the photo was
taken and direction in which the camera was pointed, and recovers the 3D coordinates of a sparse
set of points in the scene (a point cloud). While the problem of 3D reconstruction from images
has been well-studied, previous work has mainly focused on ordered sequences, such as video.
Reconstruction of Internet photo collections poses a particular challenge, as the images are given in
more or less random order, are taken with many different cameras, exhibit signicant variation in
lighting and weather, and can be extremely large in number. Thus, my pipeline is carefully designed
to robustly handle these kinds of uncontrolled input collections.
The basic principles behind recovering geometry from a set of images are fairly simple. Humans
and other animals implicitly use multi-view geometry to sense depth with binocular vision. If we
see the same point in the world (the corner of a window, say) in both eyes, we can implicitly
triangulate that point to determine its rough distance.
1
This form of depth perception depends on
two key faculties: (1) identifying parts of the images seen in the left and right eyes that correspond
to the same point in the world, and (2) knowing where the eyes are roughly located relative to each
other (to enable triangulation of corresponding points).
2
The human visual system
Similarly, given two photographs of the same (motionless) scene, a list of pixels in image A
and their corresponding pixels in image B, and the relative poses (i.e., the position and orientation)
of the cameras used to capture the images, one can calculate the 3D position of the point in the
world associated with each pair of matching pixels. To do so, we shoot a 3D ray from each camera
1
Binocular vision is only one of a number of cues we can use to estimate depth. Others include focus, parallax induced
by motion of the head, and the apparent size of objects.
2
Hence, by optically increasing this interocular distance, e.g., through a carefully engineered system of mirrors [26],
our perception of distance can be altered.
26
location through each respective matched pixel, and nd where the two rays intersect. If there is any
noise in the correspondences or the camera poses, the two rays may not intersect exactly, but we can
compute the point with the smallest distance to the two rays. The same procedure can be extended
to any number of images, as long as we know the pose for each camera, and as long as each point
we wish to triangulate is identied in at least two images.
Now suppose we want to automate this process. We rst take our digital camera and walk
around an object (a priceless Peruvian idol, say) shooting photos from different angles. Our goal is
to create a rough 3Dmodel of the object, and so we upload the photos to a computer to go through the
calculations described above. However, we run into a problem: where do we get the list of matching
points, and how do we know the vantage point from which each image was taken? Unlike the human
brain, the computer does not have an innate ability to match points between different images, nor
does it know the poses of the cameras used to take the photos. These are two fundamental problems
in computer vision.
The rst problem, that of matching 2D points between images, is known as the correspondence
problem. There are many automatic techniques for nding correspondences between two images,
but most work on the principle that the same 3D point in the world (the left eye of the idol, for
instance) will have a similar appearance in different images, particularly if those images are taken
close together.
Considering the case of two photos for a moment, suppose we can solve the correspondence
problem and identify correct pixel matches between two photos. How can we determine the poses
of the cameras? As it turns out, the correspondences place constraints on the physical conguration
of the two camera.
3
For instance, the two cameras must have been situated in such a way that rays
through corresponding pixels actually intersect (or nearly intersect, given noise in the system). This
is a powerful constraint, as two 3D rays chosen at random are very unlikely to pass close to one an-
other. Thus, given enough point matches between two images, the geometry of the system becomes
constrained enough that we can determine the two camera poses (up to a similarity transforma-
tion) [81], after which we can estimate the 3D point positions using triangulation. The problem
of using pixel correspondences to determine camera and point geometry in this manner is known
3
I refer to each image as being taken by a different camera, meaning a specic 3D pose and set of camera parameters,
even if the same physical device is used from photo to photo.
27
as structure from motion (SfM). In general, SfM methods take an arbitrary number of images and
an arbitrary number of correspondences, and estimate camera and point geometry simultaneously.
Such methods typically work by nding the conguration of cameras and 3D points which, when
related through the equations of perspective projection, best agree with the correspondences.
This chapter describes an SfMpipeline specially designed to handle diverse collections of photos
resulting from Internet search, and consists of two main stages, corresponding to the correspondence
and SfM problems. First, a set of pixel correspondences among all images is determined through
feature detection and matching. Second, an incremental SfM algorithm uses the correspondences
to estimate where the images were taken and the positions of a sparse set of 3D scene points. The
pipeline is shown as a block diagram in Figure 3.1.
The remainder of this chapter is organized as follows. Section 3.1 describes the correspondence
estimation stage, and 3.2 describes the incremental SfM algorithm. Section 3.3 discusses the time
and space complexity of these stages. Section 3.4 describes how reconstructions can be aligned to
a geocentric coordinate system, and Section 3.5 describes how recovered geometry is processed to
prepare it for viewing. Finally, Section 3.6 shows results for eleven data sets, Section 3.7 discusses
the limitations and breaking points of the system, and Section 3.8 offers concluding remarks.
3.1 Finding correspondence
The input to the correspondence estimation stage is the collection of raw images. The goal of this
stage is to nd sets of matching 2D pixel positions among the input images. Each set of matching
pixels across multiple images should correspond to a single point in 3D, i.e., each individual pixel
in a matched set should be the projection of the same 3D point. I will refer to such a corresponding
set of pixels as a point track. The name track is in analogy to tracking features across a video
sequence. Unlike with video tracking, however, where tracks are computed across sequences of
consecutive frames, tracks in the unordered setting may contain points in images widely distributed
throughout the input set.
The basic approach for nding correspondences is to identify features in different images which
have a similar appearance; two image patches that look very similar likely correspond to the same
3D point in the world. The correspondence estimation stage has three main steps: (1) nd distinctive
28
Detect features in
each image
Match keypoints
between each pair
of images
For each pair, estimate
an F-matrix and refine
matches
Chain pairwise
matches into tracks
Select good initial
image pair to seed
reconstruction
Add new images to
reconstruction and
triangulate new points
Bundle adjust
Compute correspondences
Structure-from-motion
Image correspondences
Camera and scene geometry
Input images
Figure 3.1: Block diagram of the structure from motion pipeline.
29
-0.02
-0.015
-0.01
-0.005
0
0.005
0.01
-10 -5 0 5 10
Figure 3.2: Difference of Gaussians image lter. Left: a 1D prole of the DoG lter, which is
dened as the difference between a Gaussian function and another Gaussian with somewhat smaller
width. Left: a 2D version of the lter applicable to images.
feature points in each image, (2) match features between pairs of images, and (3) link up pairwise
matches to form point tracks across multiple images.
3.1.1 Feature detection
For the feature detection step, I use the Scale-Invariant Feature Transform (SIFT) feature detec-
tor [90]. SIFT is well-suited to matching images in Internet collections due to its repeatability and
invariance to certain geometric and photometric image transformations. The SIFT detector works
by applying a differences of Gaussian (DoG) lter to the input image, then nding all local maxima
and minima of this ltered image; each (x, y) position of an extremum is kept as the location of a
feature. The DoG lter, shown in Figure 3.2, attains a maximum or minimum (or is said to re)
at the center of blob-shaped regions (e.g., the center of a dark circle on a light background, or a
light circle on a dark background). The DoG lter also res on other types of features as well, such
as corners where two thick edges meet.
However, the DoG lter only res at the center of image features whose size is roughly the
width of the lter. One of the innovations of SIFT is that it detects features at multiple scales in
the image, that is, it detects blobs of varying sizes. The result is that even if an object appears at
different scales in two imagese.g., if the Colosseum is 200 pixels wide in one image, and 400
pixels wide in anotherSIFT has the ability to match features between the two images (hence, the
30
Figure 3.3: Example set of detected SIFT features. Each detected SIFT feature is displayed as a
black box centered on the detect feature location. SIFT detects a canonical scale and orientation for
each feature, depicted by scaling and rotating each box.
scale-invariant in the name). SIFT achieves such invariance by, in essence, applying DoG lters
of varying widths to the input image, thus creating a stack of ltered images, then nding points
in this image stack that are extrema in both image space and scale space.
4
The scale (DoG radius)
at which the feature was detected is that features canonical scale. Figure 3.3 shows detected SIFT
features in a sample image, showing each features location and canonical scale. SIFT also detects
a canonical orientation for each feature by estimating a dominant local gradient direction.
In addition to detecting a set of feature positions, SIFT also computes a descriptor for each
feature, or a vector describing the local image appearance around the location of that feature (com-
puted at the features canonical scale). One simple example of a descriptor is a window of color
(or grayscale) values around the detected point. The descriptor used by SIFT considers image gra-
dients, rather than intensity values, as image derivatives are invariant to adding a constant value to
the intensity of each pixel. In fact, SIFT looks at the directions of these gradients, rather than their
4
The implementation of SIFT generates the scale space stack somewhat differently, applying a Gaussian lter to an
input image multiple times, downsampling when possible, then subtracting consecutive ltered images to create a
scale-space pyramid.
31
raw magnitude, as gradient directions are even more invariant to variations in brightness and con-
trast across images. In particular, SIFT computes histograms of local image gradient directions. It
creates a 4 4 grid of histograms around a feature point, where each histogram contains eight bins
for gradient directions, resulting in a 4 4 8 = 128-dimensional descriptor. Thus, each feature
f consists of a 2D location (f
x
, f
y
), and a descriptor vector f
d
. The canonical scale and orientation
of a feature are not used in the remainder of the pipeline. The number of SIFT features detected
in an image depends on the resolution and content of the image, but a typical 2-megapixel (e.g.,
1600 1200) image contains several thousand SIFT features.
Many other feature detectors have been developed, and some, such as MSER [92] or SURF [8]
are also designed to be invariant to common image transformations. These could also be used in an
SfM system designed for unordered collections.
3.1.2 Feature matching.
Once features have been detected in each image, the system matches features between each pair
of images. Let F(I) denote the set of features found in image I. For every pair of images I and
J, the system considers each feature f F(I) and nds its nearest neighbor (in descriptor space)
f
nn
F(J):
f
nn
= arg min
f

F(J)
[[f
d
f

d
[[
2
.
In practice, for efciency I nd an approximate nearest neighbor using the approximate nearest
neighbors library of Arya and Mount [7], which uses a kd-tree data structure to efciently compute
nearest neighbors. We now have a candidate pair of matching features (f, f
nn
). Let d
1
be the
distance between their descriptors. This candidate match is classied as a true or false match using
the ratio test described by Lowe [90]: the matching algorithm nds the second nearest neighbor,
f
nn2
F(J), to f (with distance d
2
), and accepts the match if
d
1
d
2
, the ratio of the distances to the
two nearest neighbors, is less than a threshold (I use 0.6 in practice).
After matching features in I to J, each feature f F(I) will be paired with at most one feature
in F(J). However, each feature in F(J) may be paired with many features in F(I), as a single
feature in F(J) may be closest to multiple features in F(I). The true correspondence must be one-
to-one, however, so some of these matches must be spurious. Rather than trying to reason about
32
which are correct, all such multiple matches are removed. If, after this pruning step, an image pair
has fewer than a minimum number of matches (I use sixteen), the images are deemed not to match,
and all of their feature matches are removed.
We now have a set of putative matching image pairs (I, J), and, for each matching image pair,
a set of individual feature matches. Because the matching procedure is imperfect many of these
matchesboth image matches and individual feature matcheswill often be spurious. Fortunately,
it is possible to eliminate many spurious matches using a geometric consistency test. This test
is based on the fact that, assuming a stationary scene, not all sets of matching features between
two images are physically realizable, no matter what the actual shape of the scene is. There is a
fundamental constraint between two perspective images of a static scene dened by the possible
congurations of the two cameras and their corresponding epipolar geometry.
5
The epipolar geom-
etry of a given image pair can be expressed with a 3 3, rank-2 matrix F, called the fundamental
matrix (or F-matrix), dened by the relative positions and orientations of the two cameras, as well as
internal camera settings such as zoom. Each pair of corresponding points (x, y) (x

, y

) between
two images must satisfy the epipolar constraint:
_
x

1
_
F
_

_
x
y
1
_

_
= 0 (3.1)
Thus, for a given image pair, only sets of matching features which all satisfy the epipolar con-
straint for some (initially unknown) fundamental matrix Fare admissible. We can use this fact to try
to separate the true features matches from the spurious ones: the true matches will all be explained
by the same F-matrix, while spurious matches are likely to disagree with this F-matrix (and with
each other as to what the correct F is, assuming all the false matches are independent). Finding
such an F (and a corresponding set of inliers to this F) is essentially a model-tting problem with
noisy data, a problem that can be solved using an algorithm called RANSAC (or RANdom SAmple
Consensus) [45]. RANSAC nds an F-matrix consistent with the largest number of matches by gen-
erating many different hypothesis F-matrices from random subsets of the data, counting the number
5
This constraint is very much related to the constraint between image pairs described earlier, that rays through corre-
sponding pixels must intersect.
33
of matches in agreement with each hypothesis, and keeping the best hypothesis, i.e., the one with
the most matches in agreement.
6
I run RANSACon each potential matching image pair (I, J), using the eight-point algorithm[68]
to generate each RANSAC hypothesis, and normalizing the problem to improve robustness to noise
[66]. For each hypothesized F-matrix, I use an outlier threshold of 0.6% of the maximum of the
image width and height (roughly six pixels in a 1024 768 image) to decide if each individual
feature match is consistent with that F-matrix. I then remove all outliers to the best hypothesis F-
matrix from the list of matches. If the number of remaining feature matches is less than sixteen,
I remove all of the feature matches between I and J from consideration, and deem (I, J) to be a
non-matching image pair.
Once all
_
n
2
_
image pairs have been matched, I organize the matches into point tracks by nding
connected sets of matching features across multiple images. For instance, if feature f
1
F(I
1
)
matches feature f
2
F(I
2
), and f
2
matches feature f
3
F(I
3
), these features will be grouped into
a track f
1
, f
2
, f
3
. Tracks are found by examining each feature f in each image and performing
a breadth-rst search of the set of features in other images that match f until an entire connected
component of features has been explored. These features are then grouped together into a track, and
the next unvisited feature is considered, until all features have been visited. Because of spurious
matches, inconsistencies can arise in tracks; in particular, a track can contain multiple features from
the same image, which violates the assumption that a track corresponds to a single 3D point. For
instance, if feature f
3
F(I
3
) in the example above matches a different feature f

1
,= f
1
F(I
1
),
then image I
1
will observe the track in two different locations (corresponding to features f
1
and f

1
).
Such tracks are identied as inconsistent, and any image that observes a track multiple times has all
of their features removed from that track.
Once correspondences have been found, I construct an image connectivity graph, which contains
a node for each image, and an edge between any pair of images with shared tracks. A visualization
of the connectivity graph for the Trevi data set (a collection of Internet photos of the Trevi Fountain)
is shown in Figure 3.4. To create this visualization, the graph was embedded in the plane using the
neato tool in the Graphviz graph visualization toolkit [60]. Neato works by modeling the graph as a
6
RANSAC is not specic to F-matrices, but can be used to t many other types of models to noisy data as well. Later
in this chapter RANSAC is used to t homographies and projection matrices.
34
Figure 3.4: Image connectivity graph for the Trevi Fountain. This graph contains a node (red dot)
for each image in a set of photos of the Trevi Fountain, and an edge between each pair of photos
with shared tracks. The size of a node is proportional to its degree. There are two dominant clusters
corresponding to daytime photos (e.g. image (a)) and nighttime photos (image (d)). Similar views
of the facade are clustered together in the center of the graph, while nodes in the periphery, e.g., (b)
and (c), are more unusual (often close-up) views. An interactive version of this graph can be found
at http://phototour.cs.washington.edu/imagegraphs/Trevi/.
mass-spring system and solving for an embedding whose energy is a local minimum.
The image connectivity graph for this particular collection has several distinct features. There
is a large, dense cluster in the center of the graph which consists of photos that are all fairly wide-
angle, frontal, well-lit shots of the fountain (such as image (a)). Other images, including the leaf
nodes (such as (b) and (c)) corresponding to tightly-cropped details, and nighttime images (such as
(d)), are more loosely connected to this core set. Other connectivity graphs are shown in Figures
3.10, 3.11, and 3.12, and their structure is discussed in more detail in Section 3.6.1.
3.2 Structure from motion
The second stage of the pipeline takes the set of point tracks found in the rst stage and estimates
the 3D camera and scene geometry, by nding the conguration of cameras and 3D points which,
when related through the equations of perspective projection, best agrees with the detected tracks.
This the structure from motion (SfM) problem.
35
First, let us establish some notation. The scene geometry consists of a 3D point X
j
for each
point track j. A camera can be described by its pose (position and orientation, also known as the
extrinsic parameters), and parameters modeling the internal image formation process (the intrinsic
camera parameters). I represent the extrinsic parameters of camera i with a 3-vector, c
i
, describing
the camera center, and a 3 3 rotation matrix R
i
describing the camera orientation (I will also
sometimes refer to the camera translation, t
i
, where t
i
= R
i
c
i
). Camera intrinsics are often
represented with an upper-diagonal 3 3 matrix K
i
mapping 3D rays to homogeneous 2D image
positions. In this thesis, however, I assume that this linear mapping can be modeled with a single
parameter f
i
, the focal length, and that K
i
= diag(f
i
, f
i
, 1). In general, the off-diagonal terms of the
Kmatrix model skew (the amount by which the angle between the image x-axis and y-axis deviates
from 90

) and the principal point (the point where the cameras optical axis intersects the image
plane), and the rst two diagonal entries need not be equal (the aspect ratio between the width and
height of each pixel need not be one). However, for most digital cameras, the skew is close to zero,
the principal point is near the image center, and the aspect ratio is close to one, thus the assumption
that K
i
= diag(f
i
, f
i
, 1) is reasonable [122]. While this is not an essential assumption, it reduces
the number of parameters that need to be estimated, and increases the stability of the system, as
certain parameters, especially the principal point, tend to be severely ill-conditioned. However,
certain images, such as images created by cropping a non-centered region of a large image, do not
t this model.
On the other hand, while the K matrix models a linear mapping, most consumer cameras have
noticeable non-linear lens distortion. I model this distortion with a polynomial in , the distance
from the center of the image, using a quartic polynomial

2
+
2

4
(3.2)
with two distortion parameters,
1
and
2
, for each image. I refer to the collection of camera pa-
rameters for a single camera as C
i
= c
i
, R
i
, f
i
,
1i
,
2i
. This set of parameters can be described
by nine numbers: three for the camera center c
i
, three for the rotation, as explained in Section 3.2.3,
and the three intrinsic parameters f
i
,
1i
and
2i
. Together, the camera parameters C
i
specify how a
3D point X
j
= (X
jx
, X
jy
, X
jz
) is projected to a 2D point x in image i (via the projection equation
36
Figure 3.5: Reprojection error. A 3D point X
j
is projected into camera C
i
. The reprojection error
is the distance between the projected image point, P(C
i
, X
j
) and the observed image point q
ij
.
P(C
i
, X
j
) dened below).
The point tracks found in the correspondence estimation stage can be thought of as noisy mea-
surements of the scene. In particular, each track (ideally) contains measurements of the projected
positions of a single 3D point in certain viewpoints. I denote the measured position of track j in
image i as q
ij
; note that many q
ij
are typically unknown, as not all images see all tracks. Structure
from motion is the problem of taking these measurements and jointly solving for the camera and
scene parameters that predict these measurements as well as possible. This problem is usually posed
as an optimization over the collective set of camera and scene parameters C = C
1
, C
2
, . . . , C
n

and X = X
1
, X
2
, . . . , X
m
, where an objective function measures the discrepancy between the
measured 2D point positions and those predicted by the projection equation P. For n views and m
tracks, the objective function g can be written as:
g(C, X) =
n

i=1
m

j=1
w
ij
[[q
ij
P(C
i
, X
j
)[[
2
(3.3)
where w
ij
is an indicator variable: w
ij
= 1 if camera i observes track j, and w
ij
= 0 otherwise.
The expression [[q
ij
P(C
i
, X
j
)[[ in the interior of this summation is called the reprojection error
of track j in camera i. Reprojection error is illustrated in Figure 3.5. Thus, the objective function
g is the sum of squared reprojection errors, weighted by the indicator variable. The goal of SfM
is to nd the camera and scene parameters that minimize this objective function. In my pipeline,
37
g is minimized using non-linear least squares optimization, a process known as bundle adjustment
[150].
The projection equation. The projection equation P(C
i
, X
j
) is dened as follows. First, the
point X
j
is converted to the cameras coordinate system through a rigid transformation:
X

=
_

_
X

x
X

y
X

z
_

_
= R
i
(X
j
c
i
).
Next, the perspective division is performed, and the result is scaled by the focal length:
x

=
_
_
f
i
X

x
/X

z
f
i
X

y
/X

z
_
_
.
x

=
_
x

x
, x

T
is now a 2D point in the image. Finally, radial distortion is applied to this point:

2
=
_
x

x
f
i
_
2
+
_
x

y
f
i
_
2
=
1i

2
+
2i

4
x = x

,
x is then the computed projection.
Another common way to write the projection equation is to group the camera parameters C
i
(excluding the distortion parameters) into a 3 4 projection matrix
i
:

i
= K
i
[R
i
[t
i
]
which maps homogeneous 3D points to homogeneous 2D image positions. The distortion and per-
spective division can then by applied after the projection matrix:
P(C
i
, X
j
) = r(z
div
(X
j
))
38
where z
div
performs the perspective division and r applies the radial distortion.
Because of the rotation, perspective division, and radial distortion, P is a non-linear function,
and therefore bundle adjustment is a non-linear least squares problem. Algorithms for non-linear
least squares, such as Levenberg-Marquardt, are only guaranteed to nd local minima, and large-
scale SfM problems are particularly prone to getting stuck in bad local minima. Therefore it is
important to provide good initial estimates of the parameters. Rather than trying to initialize the
parameters for all cameras and points at once, I take an incremental approach. I start with two
cameras, nd their optimal parameters, and those of the points they observe, then iteratively add one
camera at a time to the optimization.
Reconstructing the initial pair. I begin by estimating the parameters for a single pair of cameras.
This rst step is critical: if the reconstruction of the initial pair gets stuck in the wrong local mini-
mum, the optimization is unlikely to ever recover. The initial pair must therefore be chosen carefully.
The images should have a large number of matches, but also have a large baseline (distance between
camera centers), so that the initial two-frame reconstruction can be robustly estimated. I there-
fore choose the pair of images that has the largest number of matches, subject to the condition that
their matches cannot be well-modeled by a homography. A homography models the transformation
between two images of a single plane, or two images taken at the same location (but possibly with
cameras looking in different directions). Thus, if a homography cannot be t to the correspondences
between two images, it indicates that the cameras have some distance between them, and that there
is interesting 3D structure visible. These criteria are important for robust estimation of camera pose.
In particular, I estimate a homography between each pair of matching images using RANSAC
with an outlier threshold of 0.4% of max(image width, image height), and store the percentage of
feature matches that are inliers to the estimated homography. I select the initial two images to be
the pair with the lowest percentage of inliers, but which have at least 100 matches.
The system estimates the extrinsic parameters for this initial pair using Nist ers implementation
of the ve point algorithm [102],
7
and the tracks visible in the two images are triangulated, giving an
initial set of 3D points. I then perform a two-frame bundle adjustment starting from this initializa-
7
I only choose the initial pair among pairs for which a focal length estimate is available for both cameras, and therefore
a calibrated relative pose algorithm can be used.
39
Figure 3.6: Incremental structure from motion. My incremental structure from motion approach
reconstructs the scene a few cameras at a time. This sequence of images shows the Trevi data set at
three different stages during incremental reconstruction. Left: the initial two-frame reconstruction.
Middle: an intermediate stage, after fteen images have been added. Right: the nal reconstruction
with 360 photos.
tion. To perform bundle adjustment, I use the sparse bundle adjustment (SBA) package of Lourakis
and Argyros [89]. SBA is described in Section 3.2.4.
Adding new cameras and points. Next, I add another camera to the optimization. I select the
camera that observes the largest number of tracks whose 3D locations have already been estimated.
To initialize the pose of the new camera, I rst estimate its projection matrix using the direct
linear transform (DLT) technique [68] inside a RANSAC procedure. For this RANSAC step, I use
an outlier threshold of 0.4% of max(image width, image height).
Recall that the projection matrix has the form
= K[R[t] = [KR[Kt] ,
therefore the left 3 3 submatrix of (denoted
3
) is the product of an upper-triangular matrix
Kand a rotation matrix R. Kand R can therefore be computed as the RQ decomposition of
3
.
8
8
Note that, in order for it to be a valid intrinsic matrix, the upper triangular matrix K should have positive entries
along the diagonal. A unique RQ decomposition with this property is guaranteed as long as 3 is non-singular. 3
will be non-singular if the 3D points used to solve for are non-coplanar, and the corresponding 2D points are non-
colinear, which almost always is the case. Note, however, that the 3D points are often nearly co-planar, which can
cause instability in the estimation of K. In my experience, most scenes have sufcient non-planar 3D structure to avoid
this problem.
40
The translation t is then the last column of K
1
.
I use the rotation R and translation t estimated using the above procedure as the initial pose
for the new camera. In addition, I use the estimated K to either initialize the focal length of the
new camera, or verify an existing focal length estimate (see Recovering focal lengths later in this
section). Starting from this set of initial parameters, I do a round of bundle adjustment, allowing
only the new camera to change; the rest of the model is held xed.
Next, I add points observed by the new camera into the optimization. A point is added if it is
observed by at least two cameras, and if triangulating the point gives a well-conditioned estimate
of its location. I estimate the conditioning by considering all pairs of rays that could be used to
triangulate the new point, and nding the pair with the maximum angle of separation
max
. If
max
is larger than a threshold of 2

, then the point is triangulated and added to the optimization.

9
Once
the new points have been added, bundle adjustment is performed on the entire model.
This procedure of initializing a camera, triangulating points, and running bundle adjustment is
repeated, one camera at a time, until no remaining camera observes a sufcient number of points
(at least twenty in my implementation). Figure 3.6 illustrates this process for the Trevi data set. In
general only a subset of the images will be reconstructed. This subset is not selected beforehand,
but is determined by the algorithm as it adds images until no more can reliably be added.
3.2.1 Improvements to the algorithm
For increased robustness and speed, I make a few modications to the basic SfM procedure outlined
above. The rst modication deals with robustness to mismatches in the set of point tracks. Spurious
matches can have a dramatic effect on the reconstruction, pulling the recovered structure far away
from the correct solution. It is therefore critical to handle such outliers in a robust way. After
each run of bundle adjustment I detect outliers by identifying tracks that contain at least one feature
with a high reprojection error. These tracks are then removed from the optimization. The outlier
threshold for a given image adapts to the current distribution of reprojection errors for that image.
For an image I, I compute d
80
, the 80th percentile of the reprojection errors for that image, and use
9
This test tends to reject points at innity; while points at innity can be very useful for estimating accurate camera
rotations, I have observed that they can sometimes result in problems, as using noisy camera parameters to triangulate
points at innity can result in points at erroneous, nite 3D locations. See Section 3.7, Cascading errors.
41
clamp(2.4d
80
, 4, 16) as the outlier threshold (where clamp(x, a, b) clamps x to the range [a, b]).
The effect of this clamping function is that all points with a reprojection error of more than 16
pixels in any image will be rejected as outliers, and all points with a reprojection error of less than
4 pixels will be kept as inliers; the exact threshold is adaptive, and lies between these two values.
After rejecting outliers, I rerun bundle adjustment and repeat until no more outliers are detected.
The second modication is that, rather than adding a single camera at a time into the optimiza-
tion, I add multiple cameras. To select which cameras to add, I rst nd the camera with the greatest
number of matches, M, to existing 3D points, then add any camera with at least 0.75M matches
to existing 3D points. Adding multiple cameras at once results in many fewer overall iterations of
bundle adjustment and thus improved efciency.
3.2.2 Recovering focal lengths
Most digital cameras embed metadata into JPEG images in the form of Exchangeable image le
format (Exif) tags. These tags often include information on the camera make and model, the date and
time a photo was taken, and camera settings such as aperture, shutter speed, and focal length. When
an image has a focal length estimate available, it can be very helpful as initialization for the focal
length parameter of that camera in the optimization, as it can be difcult to reliably estimate focal
lengths without any calibration, particularly for planar or near-planar scenes. Exif tags invariably
provide focal length estimates in millimeters. To be useful, these estimates must be converted to
pixel units, as described in Appendix A.
I have found that most focal length estimates obtained from Exif tags are accurate enough to use
for initialization, as demonstrated in Figure 3.7; the median absolute difference between the focal
lengths calculated from Exif tags and those derived from SfM for the Trevi data set is just 3.5%.
However, the occasional image will be tagged with a wildly inaccurate focal length. Therefore,
when estimating the pose of a new camera, I check its focal length estimate against the independent
estimate obtained through the DLT procedure. Recall that the DLT procedure results in an estimate
42
0
1000
2000
3000
4000
5000
0 1000 2000 3000 4000
R
e
c
o
v
e
r
e
d

f
o
c
a
l

l
e
n
g
t
h
EXIF focal length
y = x
0
1000
2000
3000
4000
5000
0 1000 2000 3000 4000
R
e
c
o
v
e
r
e
d

f
o
c
a
l

l
e
n
g
t
h
EXIF focal length
y = x
Trevi data set Hagia Soa data set
Figure 3.7: Accuracy of Exif focal lengths. These scatter plots compare focal length estimates
obtained from the Exif tags of the image les (x-axis) with those recovered using the structure from
motion pipeline (y-axis) for two data sets (Trevi and Hagia Sophia); the line y = x is plotted for
comparison. The Exif tags were completely ignored for this experiment; initial values for focal
lengths were set using f
dlt
, and no soft constraints were used during bundle adjustment. For Trevi,
the recovered focal lengths tend to be slightly larger than those obtained from the Exif tags (by about
2.4% on average), while the average absolute difference is about 12%. However, this difference is
skewed by the presence of outliers; the median absolute difference is just 3.5%. For Hagia Sophia,
the average difference is about 0.9%, with an average absolute difference of 7.8% and a median
absolute difference of 3.1%. While in general Exif tags are accurate enough to obtain a useful
prior, the two focal length estimates occasionally vary signicantly, hence the outliers in the plot.
The reason for the estimated focal length being slightly longer may be due to a systematic source
of error. For instance, perhaps cameras use a slightly smaller region of the image sensor than the
dimensions listed in camera specications, or perhaps cameras usually round the focal length down
to the nearest discrete setting when preparing Exif tags.
of the upper-triangular intrinsic matrix K:
K =
_

_
k
11
k
12
k
13
0 k
22
k
23
0 0 1
_

_
Let f
dlt
=
1
2
(k
11
+ k
22
) be the DLT estimate of the focal length, and let f
Exif
be the focal length
calculated from the Exif tags of an image. I prefer to use f
Exif
or initialization when it exists, but
I rst test whether f
Exif
disagrees with the DLT estimate by a large factor; a large discrepancy can
indicate that f
Exif
is erroneous. In particular, I check that 0.7f
dlt
< f
Exif
< 1.4f
dlt
. If this test
43
fails, I use f
dlt
as the initial value for the focal length. Otherwise, I use f
Exif
. In addition, when
f
Exif
exists and is in agreement with f
dlt
, I add a soft constraint to the optimization to encourage
the focal length parameter f for the camera to stay close to this initial value, by adding the term
(f f
Exif
)
2
to the objective function g (using a value = 1 10
4
in all experiments). No such
constraint is used if f
dlt
is used for initialization.
3.2.3 Parameterization of rotations
During bundle adjustment, I parameterize the camera rotation with a 3-vector representing an
incremental rotation. is equal to a rotation axis (represented as a unit 3-vector n) times the angle
of rotation :
= n
and the incremental rotation matrix R(, n) is dened as:
R(, n) = I + sin [ n]

+ (1 cos ) [ n]
2

,
where
[ n]

=
_

_
0 n
z
n
y
n
z
0 n
x
n
y
n
x
0
_

_
.
The incremental rotation matrix R(, n) is pre-multiplied by the initial rotation matrix to compute
the current rotation inside the global optimization. For small incremental rotations, R(, n) is nearly
linear in .
3.2.4 Sparse bundle adjustment
To optimize the objective function g, my system uses the sparse bundle adjustment (SBA) pack-
age [89]. SBA is a non-linear optimization package that takes advantage of the special sparse struc-
ture of SfM problems to provide an algorithm with reduced time and memory complexity. At the
heart of SBA is an implementation of the Levenberg-Marquart algorithm [105], which, like many
non-linear solvers, takes as input an initial guess at the solution and nds a nearby local minimum
44
by solving a series of linear systems Hx = b (called normal equations) that locally approximate
the objective function. The matrix His an approximation to the Hessian of the objective function at
a given point, where the Hessian is a square, symmetric matrix describing the shape of a quadratic
function.
H has a row and column for each variable in the optimization. For n cameras and m point
tracks, H is a (9n + 3m) (9n + 3m) matrix; thus, H can be quite large. For 1,000 cameras
and 100,000 points (a typical problem size), Hhas 300,900 rows and columns, and nearly a billion
entries, requiring a hefty 3.6GB of memory to store as an upper-diagonal matrix of double precision
oating point numbers. The time complexity of directly solving a dense linear system is cubic in
the number of variables, also quite high.
Fortunately, however, most of the entries of Hare zeroes, and the matrix has a particular sparsity
pattern. Each entry h
ij
is non-zero if (and only if) the two variables with index i and j directly
interact (i.e., they appear together in any single reprojection error computation in the interior of g).
Hence, all nine parameters of any single camera interact, as do the three parameters of any single
point. If a given camera sees a given point, the parameters of the camera and the point interact.
However, any two distinct cameras do not directly interact, nor do any pair of distinct points; in
addition, if a camera does not observe a point, the camera and the points do not interact.
If we order the variables so that the camera parameters (grouped by camera) come rst, followed
by the point parameters (grouped by point), the sparsity pattern of H takes a form similar to that
shown in Figure 3.8. This matrix can be broken into four parts:
H =
_
_
A B
B
T
C
_
_
(3.4)
where A is the camera sub-matrix, C is the point sub-matrix, and B represents the interactions
between the cameras and points.
We can take advantage of this sparsity pattern to save memory by representing the matrix in a
sparse storage format. SBA does so, but also uses a technique called the Schur complement [150]
to improve efciency. This technique factors out the point parameters to produce a reduced camera
system whose size is the the number of camera parameters; the particular sparsity pattern of SfM
45
Figure 3.8: Sparsity pattern of the Hessian matrix for a structure from motion problem. In this im-
age, each row and column corresponds to a variable in a structure from motion problem containing
six cameras and about one hundred points. The camera parameters appear rst, followed by the
point parameters. Non-zero entries in this matrix are colored black, while zeros are colored white.
makes this reduction particularly efcient. SBA applies a direct linear solver to the reduced camera
system to obtain the new camera parameters, then back-substitutes for the new point parameters.
For large numbers of cameras and points, the bottleneck in this procedure becomes the solution of
the reduced camera system, which has cubic complexity in the number of cameras.
3.3 Algorithmic complexity
There are several parts of the SfM pipeline that require signicant computational resources: SIFT
feature detection, pairwise feature matching and F-matrix estimation, linking matches into tracks,
and incremental SfM . This section analyzes the time and space complexity of each of these steps.
I will mainly analyze time and space complexity as a function of the number of input images n.
In practice, this analysis also depends on the number of features detected in each image, which in
turns depends on factors such as image resolution and how textured each image is. However, image
resolution, and hence number of features, is bounded (each image containing a fewthousand features
on average), so I treat the number of features as roughly constant per image, thereby factoring them
out of the analysis.
46
SIFT. The feature detection step is linear in the number of input images (O(n)). Running SIFT
on an individual image, however, can take a signicant amount of time and use a large amount of
memory, especially for high-resolution images with complex textures. For instance, in the Hagia
Soa data set (described in Section 3.6) the maximum number of SIFT features detected in an image
was over 120,000. SIFT ran for 1.25 minutes on that image, and used 2.7 GB of memory, on a test
machine with a 3.80 GHz Intel Xeon processor. (The minimum number of features detected was
31; SIFT ran for 4 seconds on that image.) SIFT spent an average time of 12 seconds processing
each image in this collection (and about 5.2 hours of CPU time in total). Because SIFT runs on
each image independently, it is very easy to parallelize by assigning different sets of images to
different processors. Thus, despite being relatively time-consuming for some images, the lowoverall
complexity of the feature matching step, combined with its easy parallelization, means that it takes
a small percentage of the overall computation time.
Feature matching and F-matrix estimation. The feature matching step, on the other hand, does
take a signicant percentage of the total processing time. This is mostly due to its relatively high
complexity. Since each pair of images is considered, it has quadratic time complexity in the number
of input images (O(n
2
)). For the Hagia Soa data set, each individual image match took an average
of 0.8 seconds; in total, it took 11.7 days of CPU time to match the entire collection. Fortunately,
matching is also very easy to parallelize, as each pair of images can be processed independently.
Parallelization can compensate for the high complexity to a point, and for several example scenes
in Section 3.6, the matching stage took much more CPU time, but somewhat less wall clock time,
than the SfM stage.
Because F-matrix estimation is only run on the pairs of images that successfully match, it tends
to take a much smaller amount of time than the matching itself. Though this step is O(n
2
) in the
worst case (if all images match), for Internet collections, the percentage of image pairs that match
is usually fairly small. For instance, for the Hagia Soa collection, only 33,637 out of a possible
1,226,961 image pairs (2.7%) made it to the F-matrix estimation stage. Estimating an F-matrix for
each pair took about one hour, and 11,784 of the 33,637 pairs survived this step. This step is also
easily parallelized.
47
Linking up matches to form tracks. To implement this step, I perform a breadth-rst search on
the graph of feature matches, marking each feature as it is visited. Each feature in each image is
visited once (i.e., O(n) features are visited), and each individual feature match is also visited once
(in the worst case, O(n
2
) matches are visited). However, the graph of feature matches also tends
to be quite sparse. For the Hagia Soa collection, there were 1,149,369 individual feature matches
left after the F-matrix estimation step, about 100 per matching image. Grouping matches into tracks
took a total of just six minutes for this collection.
Structure from motion. The main computational bottleneck in the incremental structure from
motion algorithm is the bundle adjustment stage computed using SBA.
10
As described in Sec-
tion 3.2.4, SBA uses the Levenberg-Marquardt (LM) algorithm to nd a local minimum of the
objective function. Each iteration of LM involves solving a linear system; in SBAs implementation
this is done using a dense, direct method such as LU factorization. Solving the reduced camera
system using such methods takes time cubic ((t
3
)) in the number of cameras t in the current re-
construction [56]. Each call to SBA is capped at a maximum number of iterations, thus the total run
time of each call to SBA is still (t
3
). During the course of the SfM stage, SBA is called repeatedly
with more and more images as the reconstruction grows. The total running time of SfM thus de-
pends on how many times SBA is called, and on how many images are added to the reconstruction
during each iteration of SfM.
If all the images are added immediately after the initial two-frame reconstruction, there will be a
single call to SBA with all n images, and the total running time for SfM will be (n
3
). This (n
3
)
complexity will also hold if SBA is called a constant k number of times, e.g., if on average some
proportion
n
k
of the input images are added during each iteration of SfM (see Appendix B for the
details of this analysis). If, on the other hand, a roughly constant number of images are added per
iteration, independent of n, then the total running time will be (n
4
). Roughly speaking, each call
to SBA with t images does work in proportion to t
3
, so when images are added per iteration, the
10
Time is also spent in SBA triangulating new points and estimating the pose of new cameras, but the running times of
these steps are dominated by bundle adjustment.
48
total amount of work done by SBA over the course of SfM grows roughly as:
n/

i=1
(i )
3
=
3
n/

i=1
i
3
=
3
_
_
n

_ _
n

+ 1
_
2
_
2
= O(n
4
)
Appendix B presents a more detailed analysis showing that the running time of this scenario is
(n
4
).
The total running time depends on the properties of the collection. If all the photos are very
similar, they may be added in large batches, and the running time may be close to (n
3
). If, on
the other hand, the images are taken with a roughly uniform sampling along some path, it is likely
that a constant number of images will be added at each iteration (on average), resulting in a total
running time of (n
4
). In my experience, Internet photo collections behave more like the latter
case, resulting in run times closer to (n
4
).
11
In the future, analyzing the behavior of the algorithm
on a larger collection of image sets will allow for better characterization of the (empirical) run time.
In addition to the high complexity of the SfM stage, out of all the stages described in this section,
SfM is the most difcult to parallelize. Doing so would require either a parallel implementation of
a linear system solver, or a way to subdivide the images into independent groups. Despite these
difculties, in practice the time required for SfM is often less than the time spent matching images.
For instance, SfM took only 5.5 hours for the Hagia Soa data set, compared to more than one
day (of wall clock time) for matching. There may be several reasons for this. First, typically only
a subset of the images in a collection are ultimately registered. For instance, in the Hagia Soa
collection, 446 of the 1567 input images (28%) were registered. Second, I use the highly optimized
Intel Math Kernel Library (MKL) linear solver [70]. For an example 600 image data set, MKL
solves the reduced camera system in about 27 seconds. This is a signicant amount of time, but for
collections of 1,000 photos or less this typically leads to running times on the order of days (rather
than weeks). Nevertheless, for truly massive, well-connected collections, SfM would likely dwarf
the amount of time spent in the rest of the pipeline.
SBA uses O(nm) amount of memory (for n cameras and m points) to store the matrix of binary
indicator variables w
ij
dening which points are visible in which views, and O(n
2
) memory to store
11
This despite the fact that only a subset of images are ultimately registered, and SfM is only run on that subset. The
size of this subset tends to grow in proportion to the total number of images, however.
49
the reduced camera matrix. In practice, the matrix of indicator variables is quite sparse, so a more
memory-efcient sparse matrix representation could be used. The reduced camera matrix is also of-
ten sparse (an entry corresponding to two cameras is only non-zero if the cameras have shared point
tracks), and thus could also be potentially more storage-efcient in practice. In addition, reordering
techniques [150] could be used to reduce ll-in while factoring the reduced camera matrix.
3.4 Geo-registration
The SfM procedure above estimates relative camera locations, reconstructing geometry up to an
unknown similarity transformation. These relative camera locations are sufcient for many tasks,
including most of the scene visualization applications described in Part II, which do not absolutely
require knowing where in the world a scene exists, or how large the scene is. However, other appli-
cations require more information. For instance, if we want to measure distances in the reconstructed
scene in meaningful units, we need to know the scale of the reconstruction. If we want to display
the reconstruction on a site such as Google Maps, we need to know the latitude and longitude of
each camera. For these tasks, I describe techniques for aligning the recovered model with a geo-
referenced image or map (such as a satellite image, oor plan, or digital elevation map) so as to
determine the absolute geocentric coordinates of each camera.
Given a perfectly accurate reconstruction, the estimated camera locations are related to the ab-
solute locations by a similarity transform (global translation, 3D rotation, and uniform scale). My
system provides a simple manual interface for determining the alignment of a model to a geolocated
map or overhead image; the user interactively rotates, translates, and scales the model until it is in
agreement with a provided map. To assist the user, I rst estimate the up or gravity vector, as
described in Section 3.5. The 3D points and camera locations are then rendered superimposed on
the alignment image, using an orthographic projection with the camera positioned above the scene,
pointed downward. If the up vector was estimated correctly, the user needs only to rotate the model
in 2D, rather than 3D. In my experience, it is usually fairly easy, especially in urban scenes, to per-
form this alignment by matching the recovered points to features, such as building facades, visible
in the image. Figure 3.9 shows a screenshot of such an alignment.
In some cases the recovered scene cannot be well-aligned to a geo-referenced coordinate system
50
Figure 3.9: Example registration of cameras to an overhead map. Here, the cameras and recovered
line segments from the Prague data set are shown superimposed on an aerial image. (Aerial image
shown here and in Figure 5.1 courtesy of Gefos, a.s. [51] and Atlas.cz.)
using a similarity transform. This can happen because errors accumulated during reconstruction
can cause low-frequency drift in the recovered point and camera locations. Drift does not have a
signicant effect on the visualization tools described in Part II, as the error is not usually locally
noticeable, but is problematic when an accurate model is desired.
One way to straighten out the recovered scene is to pin down a sparse set of ground control
points or cameras to known 3D locations (acquired, for instance, from GPS tags attached to a few
images) by adding constraints to the SfM optimization. Alternatively, a user can manually specify
correspondences between points or cameras and locations in an image or map, as in the work of
Robertson and Cipolla [120].
3.4.1 Aligning to Digital Elevation Maps
For landscapes and other very large scale scenes, we can take advantage of Digital Elevation Maps
(DEMs), used for example in Google Earth [57] and with coverage of most of the United States
available through the U.S. Geological Survey [151]. To align point cloud reconstructions to DEMs,
I manually specify a few correspondences between the point cloud and the DEM, and estimate a 3D
similarity transformto determine an initial alignment. I then re-run the SfMoptimization with an ad-
51
ditional objective termto t the specied DEMpoints. In the future, as more geo-referenced ground-
based imagery becomes available (e.g., through websites like WWMX [149], Photosynth [109], or
Flickr), this manual step will no longer be necessary.
3.5 Processing the recovered scene
The output of the reconstruction pipeline is organized into a scene description with the following
elements:
A set of points P = p
1
, p
2
, . . . , p
n
. Each point p
j
has a 3D location, X
j
, and a color,
Color(p
j
), obtained by projecting the point into one of the images that observes it and sam-
pling the color at that image point.
A set of cameras parameters, C = C
1
, C
2
, . . . , C
k
.
A visibility mapping, Points(), between cameras and the points they observe. Points(C) is
the subset of P containing the points observed by camera C. Points(C) is derived from the
set of point tracks created in the correspondence estimation stage, and not from an analysis
of occlusions (which would require dense geometry). Therefore, the visibility mapping is
approximate. A point p Points(C) is very likely to be visible to camera C, whereas a
point q / Points(C) is possibly visible to C. q may be outside the eld of view or occluded
(and thus correctly excluded from Points(C)), or it may have not been correctly detected and
matched (and incorrectly excluded).
After a scene is reconstructed, a few additional processing steps are required to clean the geom-
etry and prepare the reconstructed scene description for viewing. These steps are:
1. Remove low-condence points and cameras.
2. Remove radial distortion from the photos.
3. Estimate the up vector for the reconstruction.
4. Compute a set of proxy planes for use in scene rendering.
5. Optionally, detect 3D line segments.
52
Removing low-condence points and cameras. Occasionally, due to bad matches or weak ge-
ometry, spurious points and misregistered cameras can appear in a reconstruction. The rst pro-
cessing step attempts to automatically detect and remove these undesirable points and cameras by
identifying low-condence geometry. A point p is deemed low-condence if:
1. p is seen by fewer than three cameras; a point seen by only two views is much more likely to
be spurious than one seen in three or more views, as the likelihood of nding a spurious point
that is geometrically consistent with k views dramatically decreases as k grows.
2. The maximum angle over all pairs of rays that see p is less than a threshold
max
(points
which are seen by views which are too close together can have very high uncertainty). I use

max
= 1.5

.
Points which meet either of these criteria are pruned from the point set P. After points have been
pruned, images which now see fewer than sixteen points are removed as low-condence views. This
processing step often cleans up the geometry considerably.
Estimating the up vector. As described earlier, SfM can only recover geometry up to an unknown
similarity transform, and thus the recovered scene can be in an arbitrary orientation; in particular,
there is no natural up direction (often called the up vector in computer graphics). Thus, to make
sure the scene is properly oriented, I estimate the up vector using the method of Szeliski [144]. This
method uses the observation that most photos are taken without signicant twist (rotation about the
cameras viewing direction), and computes the up vector which minimizes the amount of residual
twist over all cameras. This observation is largely correct, with the exception that cameras are
sometimes rotated by 90

(for portrait shots); such photos can appear sideways in image viewing
software if the camera fails to properly orient them at the time they are taken. While such images can
skew the up vector computation, they are also usually easy to identify, because in most collections
the majority of photos are property oriented, and the remainder appear as outliers.
Removing radial distortion. For each registered photo I
j
, a new, undistorted, image is created
by applying the inverse distortion transform to I
j
. It is not straightforward to compute the inverse
transform in closed form from the distortion parameters
1j
and
2j
, so I compute an approximate
53
inverse transform by tting a degree-4 polynomial to sampled pairs of distorted and undistorted
points.
Computing proxy planes for rendering. The visualization tools described in Part II set up virtual
projection planes in the scene onto which images are projected for rendering. These planes should
match the actual scene geometry (e.g., coincide with the sides of buildings) as well as possible.
The tools in Part II require a projection plane for each camera (for scene rendering), as well as a
projection plane for each pair of cameras (for creating transitions between two images):
For each camera C
i
, I compute a 3D plane, Plane(C
i
), by using RANSAC to robustly t a
plane to Points(C
i
).
For each pair of neighboring cameras C
i
and C
j
(cameras which view at least three points in
common), I compute a 3D plane, CommonPlane(C
i
, C
j
) by using RANSAC to t a plane to
Points(C
i
) Points(C
j
).
Many planar surfaces which occur in the world are vertical (walls) or horizontal (ground and ceiling
planes). However, few images are taken looking directly at the ground or other horizontal surfaces,
so I normally constrain Plane(C
i
) and CommonPlane(C
i
, C
j
) to be vertical. This constraint often
helps produce better renderings in the visualization tools. To enforce this constraint, I project all the
points onto the ground plane, and t a 2D line, rather than a 3D plane, to the projected points.
Detecting line segments. Finally, I detect 3D line segments using a method similar to that of
Schmid and Zisserman [126]. This technique is described in Appendix C.
The scene representation, originally consisting of the point set P, the camera set C, and the
visibility mapping, Points, is now augmented with:
A set of 3D line segments L = l
1
, l
2
, . . . , l
m
and a mapping, Lines, between cameras and
sets of visible lines.
The set of computed planes, Plane(C
i
) and CommonPlane(C
i
, C
j
).
54
3.6 Results
I have applied my system to many different input photo collections, including uncontrolled Internet
sets consisting of images downloaded from Flickr, and collections captured by a single person. In
each case, my system detected and matched features on the entire set of photos and automatically
identied and registered a subset corresponding to one connected component of the scene. This
section gives results for several different scenes, including eight Flickr collections:
1. Trevi Fountain, a set of photos of the Trevi Fountain in Rome.
2. St. Basils, photos of Saint Basils Cathedral in Moscow.
3. Mount Rushmore, photos of Mount Rushmore National Monument, South Dakota.
4. Sphinx, photos of the Great Sphinx of Giza, Egypt.
5. Hagia Sophia, photos of the Hagia Sophia in Istanbul.
6. Yosemite, photos of Half Dome in Yosemite National Park.
7. Colosseum, photos of the Colosseum in Rome.
8. Notre Dame, photos of the Notre Dame Cathedral in Paris.
Three other sets were taken in more controlled settings (i.e., a single person with a single camera
and lens):
1. Great Wall, a set of photos taken along the Great Wall of China.
2. Prague, photos of the Old Town Square in Prague.
3. Annecy, photos of a street in Annecy, France.
More information about these data sets (including the number of input photos, number of regis-
tered photos, CPU time, and average reprojection error), is shown in Table 3.1. The running times
reported in this table were generated by running the complete pipeline on a 3.80GHz Intel Xeon
machine with 4GB of core memory. While total CPU time is reported in Table 3.1, in practice,
the keypoint detection and matching phases were run in parallel on ten such machines. Table 3.2
contains information on how much time was spent in each stage of the reconstruction pipeline for
each data set. While the majority of CPU cycles were spent on image matching, this phase is much
55
Table 3.1: Data sets and running times. Each row lists information about each data set used: Collec-
tion, the name of the set; Search term, the Flickr search term used to gather the images; # photos,
the number of photos in the input set; # reg. the number of photos registered; # points, the number
of points in the nal reconstruction; CPU time, the total CPU time for reconstruction; error, the
mean reprojection error, in pixels, after optimization. The rst eight data sets were gathered from
the Internet, and the last three were each captured by a single person. In each subset, the rows are
sorted by number of input views.
Collection Search term # images # reg. # points CPU time error
Trevi Fountain trevi rome 466 370 114742 2.8 days 0.698
St. Basils basil red square 627 197 25782 3.8 days 0.816
Mt. Rushmore mount rushmore 1000 437 131908 10.2 days 0.444
Sphinx sphinx egypt 1000 511 130182 10.2 days 0.418
Hagia Sophia hagia sophia 1567 446 82880 12.2 days 0.395
Yosemite halfdome yosemite 1882 678 264743 23.8 days 0.757
Colosseum colosseum (rome roma) 1994 964 425828 23.3 days 1.360
Notre Dame notredame paris 2635 598 305535 36.7 days 0.616
Great Wall N/A 120 81 24225 0.3 days 0.707
Prague N/A 197 171 38921 0.5 days 0.731
Annecy N/A 462 420 196443 3.5 days 0.810
Table 3.2: Running times for each stage of the structure from motion pipeline. This table breaks out
the total CPU time into time spent (a) detecting features in each image (Feat. extract), (b) matching
features between each pair of images (Matching), and (c) structure from motion (SfM). The SfM
stage was run on a single machine with a 3.8GHz Intel Xeon processor, and the feature detection
and matching stages were run in parallel on ten such machines. The overall wall clock time spent
during reconstruction is shown in the last column.
Collection # images # reg. Feat. extract Matching SfM Total CPU Total wall
Trevi Fountain 466 370 1.6 hrs 2.0 days 15.6 hrs 2.8 days 20.7 hrs
St. Basils 627 197 2.1 hrs 3.7 days 1.2 hrs 3.8 days 10.3 hrs
Mt. Rushmore 1000 437 3.4 hrs 9.4 days 15.4 hrs 10.2 days 1.6 days
Sphinx 1000 511 3.4 hrs 9.4 days 15.7 hrs 10.2 days 1.6 days
Hagia Sophia 1567 446 5.2 hrs 11.7 days 6.6 hrs 12.2 days 1.5 days
Yosemite 1882 678 6.4 hrs 19.4 days 4.1 days 23.8 days 6.3 days
Colosseum 1994 964 6.8 hrs 18.9 days 4.1 days 23.3 days 6.3 days
Notre Dame 2635 598 9.0 hrs 33.1 days 3.2 days 36.7 days 12.7 days
Great Wall 120 81 0.3 hrs 3.2 hrs 2.5 hrs 0.3 days 2.8 hrs
Prague 197 171 0.4 hrs 8.7 hrs 2.2 hrs 0.5 days 3.1 hrs
Annecy 462 420 1.6 hrs 1.7 days 1.8 days 3.5 days 1.9 days
56
easier to parallelize than SfM, and in practice matching (on 10 machines) often took less wall clock
time than SfM. Visualizations of these data sets are shown in Figures 3.10, 3.11, and 3.12.
These results demonstrate that the reconstruction system can robustly handle a variety of scenes:
indoor scenes such as the Hagia Soa, natural scenes such as Half Dome and Mount Rushmore, and
urban scenes such as Annecy. However, these are just a small sample of the more than one hundred
collections that have been successfully reconstructed with my system. My system has also been
used by other researchers. Examples of reconstructions used in other systems, and the applications
they have helped enable, can be found in [2], [10], and [55]. The system has also been used to aid
in reconstruction of over twenty different scenes in Florence, Italy, for an installation by the artist
Marnix de Nijs [31], and is the original basis for Microsofts Photosynth [109], a online system for
reconstructing and viewing scenes that has been applied to thousands of photo collections.
3.6.1 Discussion
The image connectivity graphs shown in the third column of Figures 3.10-3.12 suggest that there is
a common pattern among Internet photo collections: they consist of several large clusters of photos,
a small number of connections spanning clusters, and a sparse set of leaves loosely connected to the
main clusters. The large clusters usually correspond to sets of photos from similar viewpoints. For
instance, the large cluster that dominates the Mount Rushmore connectivity graph are all images
taken from the observation terrace or the trails around it, and the two large clusters on the right side
of the Colosseum connectivity graph correspond to the inside and the outside of the Colosseum.
Sometimes clusters correspond not to viewpoint but to different lighting conditions, as in the case
of the Trevi Fountain collection (see Figure 3.4), where there is a daytime cluster, a nighttime
cluster, and a sparser set of links connecting the two.
The neato mass-spring system, used to embed the graph into the plane, acts in a way similar
to multidimensional scaling: similar images will tend to be pulled together, and dissimilar images
are pushed apart. This behavior can result in photos being laid out along intuitive dimensions. For
instance, in the large daytime cluster at the top of the connectivity graph for the Trevi dataset, as
well as the sparser cluster on the bottom, the x-axis roughly corresponds to the angle from which
the fountain is viewed.
57
sample image reconstruction graph (matches) graph (reconstructed)
Figure 3.10: Sample reconstructed Internet photo collections. From top to bottom: Trevi, St.
Basils, Rushmore, and Sphinx. The rst column shows a sample image, and the second col-
umn shows a view of the reconstruction. The third and fourth columns show photo connectivity
graphs, in which each image in the set is a node and an edge links each pair of images with feature
matches. The third column shows the photo connectivity graph for the full image set, and the fourth
for the subset of photos that were ultimately reconstructed.
58
sample image reconstruction graph (matches) graph (reconstructed)
Figure 3.11: More sample reconstructed Internet photo collections. From top to bottom: Hagia
Soa, Yosemite, Colosseum, and Notre Dame.
59
Figure 3.12: Sample reconstructed personal photo collections. From top to bottom: Great Wall,
Prague, and Annecy. Each of these three photo collections were taken by a single person.
60
While the graphs in the third column of Figures 3.10-3.12 represent the connectivity of entire
photo sets, the fourth column shows the subgraph the SfM algorithm was able to reconstruct. As
described in Section 3.2, the algorithm does not, in general, reconstruct all input photos, because
the input set may form separate connected components, or clusters that are too weakly connected
to be reliable reconstructed together. These subgraphs suggest that for unstructured datasets, the
reconstruction algorithm tends to register most of one of the main clusters, and can sometimes
bridge gaps between clusters with enough connections between them. For instance, in the Sphinx
collection, the SfM algorithm reconstructed two prominent clusters, one on the right side of the
graph, and one on the bottom. These clusters correspond to two sides of the Sphinx (the front and the
right side) which are commonly photographed; a few photos were taken from intermediate angles,
allowing the the two clusters to be connected. Similarly, in the Colosseum collection, two weakly
connected clusters corresponding to the inside and outside of the Colosseum were reconstructed. In
general, the more connected the image graph, the greater the number of images that can successfully
be registered.
For the controlled datasets (Annecy, Prague, and Great Wall), the photos were captured with
the intention of generating a reconstruction from them, and thus the images are more uniformly
sampled; as a result, the connectivity graphs are less clustered. In the Prague photo set, for instance,
most of the photos were taken all around the Old Town Square, looking outward at the buildings. A
few were taken looking across the square, so a few longer range connections between parts of the
graph are evident. The SfM algorithm was able to register most of the photos in these datasets.
3.7 Failure modes
While a large percentage of the time the SfM algorithm produces qualitatively good reconstructions,
it sometimes fails, resulting in bad, or no, geometry. I have observed four main failure modes:
1. Insufcient overlap or texture. The input images are simply too sparse, or too textureless,
for the system to nd sufcient correspondences between the images.
2. Ambiguous, repeating textures. Many scenes, especially urban environments, exhibit repet-
itive textures or symmetry. While my algorithm does well on such scenes surprisingly often,
there are times when two parts of a scene are simply too similar, and are confused as identical
61
objects.
3. Bad initialization. As described in Section 3.2, the reconstruction of the initial image pair is
critical; a bad start is almost impossible to recover from.
4. Cascading errors. Occasionally, the algorithm will make a mistake in the placement of a
camera. Such errors can propagate and be compounded further along the pipeline as points
and other images that depend on that view are added, leading to misestimation of large sec-
tions of a reconstruction.
I now discuss each of these failure modes, in turn.
Insufcient overlap or texture. To ensure enough overlap for reconstruction, a general rule of
thumb is that every point in the scene must be visible, and must be matched, in at least three images
(the rule of three [110]). Consider, for instance, a sequence of photos I
1
, I
2
, . . . , I
n
of a row of
at building facades, taken by a person walking down a city street. The rule of three implies that
pairs of images taken two apart, such as I
1
and I
3
, must have some overlap, otherwise no point
would be visible in more than two images. Thus, each pair of neighboring images, such as I
1
and
I
2
must have more than 50% overlap.
Even when points are visible in enough images, they may not be linked up correctly due to
limitations in feature detection and matching. For instance, my algorithm cannot reconstruct a
complete scene from the images in the nskulla multi-view stereo data set, shown in Figure 3.13.
12
Three representative neighboring images (labeled A, B, and C) are shown in the gure; feature
matches were found between A and B, as well as between B and C, but none were found between
A and C. These two images are taken too far apart; their rotation angle about the skull is too
large. Hence, there is no feature that was seen by A, B, and C at oncethe rule of three is not
satised. This is true of many triplets in the collection, and this collection could not be reconstructed.
Interestingly, some multi-view stereo algorithms do well on this collection [54, 49].
Ambiguous, repeating textures. Many buildings exhibit a regularity that can confuse SfM. For
instance, the Florence Duomo, shown in Figure 3.14, has an overall shape which is largely symmet-
12
The nskulla set is available at Yasutaka Furukawas collection of data sets at http://www-cvr.ai.uiuc.
edu/ponce_grp/data/mview/.
62
Image A Image B Image C
Figure 3.13: Three neighboring images from the nskulla data set. The structure from motion algo-
rithm fails on this data set due to too little overlap between views. Images A and B have sufcient
matches, as do B and C, but A and C have too few. Having matches between triplets of views is a
requirement for the SfM algorithm.
Figure 3.14: Incorrect interpretation of the Florence Duomo. Left: an overhead viewof the Florence
Duomo. Right: a reconstruction of the Florence Duomo from a collection of 460 images. The
Duomo is mistakenly unwrapped, with the north and south sides of the cathedral joined together
on the south side. That is, the north wall is erroneously reconstructed as an extension of the south
wall, and the west wall appears again on the east side of the reconstruction; these problem areas are
highlighted in red. This problem occurs because both sides of the dome area look very similar, and
are confused as the same side by the SfM algorithm.
63
Figure 3.15: Erroneous matches between images of the Arc de Triomphe. The two images shown
are of opposite sides of the monument (note the differences in the large sculptures on the right side
of the images). Yet many features are extremely similar; lines are drawn between points mistakenly
identied as matches during correspondence estimation. The SfM algorithm merges the two sides
into one.
ric about its east-west axis, a dome which is symmetric about multiple axes, and a uniform striped
texture on all sides. The reconstructed Duomo appears to be nearly twice as long as the real Duomo.
What happened is that the matching algorithm matched images of the north side of the dome with
images of the south side of the dome; thus both sides of the dome become the same sideit is as if
the north side of the Duomo were picked up, rotated 180 degrees about the center of the dome, and
set back down. The north wall then extends out towards the right, and the west wall is mirrored on
the east side of the building.
Another example of erroneous matching is shown in Figure 3.15. As the gure shows, the two
sides of the Arc de Triomphe are quite similar, though not identical, and images of different sides
can occasionally have enough feature matches to be deemed matching images. My SfM algorithm
thus merges the two sides into one.
Bad initialization. Another source of error is erroneous initialization, due to the sensitivity of the
optimization to the reconstruction of the initial pair. There are two common ways the initialization
can fail:
1. Necker reversal. This problem is a generalized version of the well-known Necker cube illu-
sion in which a line drawing of a cube has two possible 3D interpretations. The same ambi-
64
Figure 3.16: Example of Necker reversal on the dinosaur sequence. Left: two images from the
data set. Middle: the correct interpretation of the scene. The images are taken above the dinosaur,
looking down. Right: Necker-reversed solution. The images are reconstructed below the dinosaur,
looking up. The Necker-reversed dinosaur is (approximately) a reection of the correct solution.
[Images used with thanks to Wolfgang Niem, University of Hanover, and the Oxford Visual Geom-
etry Group.]
guity arises in multiple orthographic views of a 3D scene; two different, stable interpretations
of the shape of the scene and positions of the cameras exist, hence there is an ambiguity in
the reconstruction. For perspective images, the symmetry of the two solutions is broken; the
correct interpretation will generally be the global minimum of the SfM objection function, but
the Necker-reversed solution may still be a strong, stable local minimum. If the initial image
pair is reconstructed with the incorrect, Necker-reversed solution, it is exceedingly likely that
the nal solution will also be reversed. An example of a Necker-reversed reconstruction (of
the Oxford dinosaur sequence) is shown in Figure 3.16. When Necker reversal occurs, choos-
ing a different starting pair generally solves the problem; alternatively, the user can explicitly
tell the system to reverse the initial solution. Brown presents a different possible solution:
trying out both interpretations for the initial pair and choosing the two-frame reconstruction
with the lower reprojection error [15] to seed the rest of the reconstruction procedure.
2. Insufcient parallax. On occasion, the algorithm can pick a pair of images that have insuf-
cient baseline to admit a well-conditioned reconstruction. Recall that to estimate the amount
of parallax between an image pair, I t a homography to the feature matches, and compute the
percentage of points that are outliers. This estimate of parallax can sometimes be inaccurate;
65
Figure 3.17: Cascading errors in the Colosseum. This overhead view of a reconstruction of the
Colosseum shows evidence of a large problem. The interior is well-reconstructed, but the outer wall
(the spray of points curving towards the top of the image) is jutting out from the interior at nearly
right angles. This problem occurred because the inside and outside of the Colosseum are very
weakly connected. The reconstruction began from the inside, and proceeded correctly until images
of the outside began being added; at that point a few mistakes occurred and were compounded as
the reconstruction of the outside grew. Interestingly, when the reconstruction is started from images
of the outside of the Colosseum, this problem does not occur (see the reconstruction in Figure 3.11),
suggesting that the order in which the images are added can make a difference.
for example, if a pair of images are taken close together (and therefore should be explained
by a homography), but there are a large number of spurious matches, many of these spurious
matches may be counted as outliers. Again, manually selecting a different starting pair can
often solve this problem.
Cascading errors. When the algorithm makes a mistake in placing a camera, that error can propa-
gate to later stages of the algorithm; points observed by that view can be triangulated wrongly, views
which are in turn initialized by erroneous points will be misestimated, and so on. An example of
this problem is shown in Figure 3.17, in which part of the Colosseum has been reconstructed quite
poorly (part of the outside wall is jutting out from the rest of the Colosseum at nearly right angles).
The most common reasons for this type of problem are twofold:
66
1. Bad camera initialization. If a new camera views a relatively small number of 3D points
and enough of them are outliers, then the RANSAC algorithm for nding the initial pose of
the camera can fail, resulting in an initial estimate with signicant error.
2. Large uncertainty. Related to the initialization problem is the problem of large uncertainty
in parameter estimates. Under certain circumstances, a camera can be initialized more or less
correctly, but will have large uncertainty. For instance, a new camera which views a set of
existing 3D points all clustered together in a small region of the image may not be estimated
with high certainty. As another example, consider a view whose pose is estimated using only
relatively distant points (on a building 100 meters away, say). Relative to the distance to
that building, the uncertainty in the position of this camera might be lowperhaps one or
two metersbut relative to much closer objects the uncertainty can be quite signicant. If
the view is used to triangulate points on an object three meters away, the points may have
signicant error (despite having low reprojection error). These problems can potentially be
avoided if the system were extended to model uncertainty in parameter estimates.
Once a problem of this sort occurs, it can result in the reconstruction breaking apart into two irrec-
oncilable pieces, one built before the mistake, and one after, as in the case of Figure 3.17, where the
inner and outer parts of the Colosseum have been reconstructed in seemingly unrelated coordinate
systems.
3.8 Conclusion
This chapter presented a structure from motion pipeline adapted to diverse collections of unordered
images, demonstrated results on several data sets, and discussed common failure modes. The main
contribution is the pipeline itself, the rst to be demonstrated on large collections of photos obtained
from Internet search.
Scale is still a primary challenge with structure from motion. Several reconstructions described
in Section 3.6 took several days to compute; the Notre Dame reconstruction took nearly two weeks.
Most of the CPU cycles are concentrated in two parts of the pipeline: the pairwise matching stage
and the incremental SfM stage. Several researchers have investigated linear time methods for image
matching using bag-of-words models [104, 24] to help address the matching problem. To improve
67
the efciency of SfM , several existing techniques could be applied, such as the work of Steedly, et
al., on applying spectral partitioning to SfM to reduce the number of parameters.
However, one signicant aspect of the scale problem for Internet photo collections is that such
collections are often highly redundant, with similar images appearing over and over again. Redun-
dant views contribute little additional information to the reconstruction, but weigh the optimization
down with extra parameters. In addition, such views lead to clusters of point tracks that are visi-
ble in many images, increasing the density of the reduced camera system. This increased density
tends to slow down the non-linear optimization engine even further. The next chapter introduces a
technique for reducing the amount of redundancy in a collection by intelligently selecting a subset
of images to reconstruct. This technique can signicantly improve the efciency of SfM on large
photo collections.
68
Chapter 4
SKELETAL GRAPHS FOR EFFICIENT STRUCTURE FROM MOTION
The basic structure from motion pipeline presented in the previous chapter demonstrates the
possibility of adapting structure from motion (SfM) methods, originally developed for video, to
operate successfully on unstructured Internet photo collections. The logical nal step is to run this
algorithm on the ultimate unstructured collectionall the images on the Internet. Unfortunately, the
reconstruction algorithm of the last chapterand others in the current generation of unstructured
SfM methods [124, 15, 153]simply do not scale to the thousands or tens of thousands of images
typical of image search results, let alone the billions of images on the entire Internet. Taking just
one example from Chapter 3, a single building, the Notre Dame cathedral, took nearly two weeks
to reconstruct from 2,600 photos, in an optimization involving over 300,000 points and nearly a
million parameters; a challenging task, and for just one building! While the pairwise matching
stage took a signicant fraction of this time, linear-time image matching algorithms, based on ideas
from information retrieval, are beginning to appear [104, 24]. Thus, this chapter will focus on
the SfM stage. While there exist techniques for scalable SfM for video, such as sub-sampling
and hierarchical decomposition [46, 101], these are not directly applicable to Internet collections.
Internet collections have very different properties from video: not only they unordered, but they also
tend to be highly oversampled in some regions of popular viewpoints and undersampled in others,
as can be observed in some of the image connectivity graphs shown in the previous chapter.
Intuitively, however, the difculty of reconstructing a scene should depend on the complexity
of the scene itself, not the number of images. For large Internet photo collections a much smaller
subset of images may be sufcient to represent most of the information about the scene. If we
could identify such a subset of views, we could focus our reconstruction effort on these images and
produce truly scalable algorithms.
The key technical problem is to identify a subset of views that maximizes the accuracy and com-
pleteness of a reconstruction while minimizing the computation time. This is of course a tradeoff:
69
(a) (b) (c)
Figure 4.1: Image graph and skeletal graph for the Stonehenge data set. (a) A few images from the
Stonehenge data set. (b) The image graph G for the full set of images from this data set. The graph
forms a large loop due to the circular shape of Stonehenge, but is more densely sampled in some
regions than others. (c) A skeletal graph for G. This graph retains the same overall topology, but is
much sparser.
the more views we leave out, the faster we can reconstruct the scene, but the less complete and
accurate the reconstruction is likely to be.
Even expressing this tradeoff by translating high-level concepts like accuracy, completeness and
run time into mathematical objectives that can be optimized in the context of SfM is a challenging
problem. However, we can get an intuitive sense of this tradeoff by examining the image connectiv-
ity graphs in Figure 4.1. This gure shows a connectivity graph for a set of images of Stonehenge;
the full graph, not surprisingly, forms a ring, since people photograph Stonehenge from all sides.
This graph is quite dense in some regions (e.g., near the top of the graph), but much less dense in
others (such as the region near ve oclock). Now consider the graph on the right (which I refer
to as a skeletal graph). This graph is much sparser than the full graph, but still retains the overall
circular topology of the full graph and (nearly) spans the entire image collection; in some sense, it
captures most of the information in the full graph, with many fewer edges.
To turn this intuition into a concrete optimization problem, I make two approximations. First, I
optimize uncertainty instead of accuracy, since the former can be computed without knowledge of
ground truth geometry. Second, I use the number of images as a proxy for run time, which enables
an algorithm-independent measure of efciency. Completeness is ensured via the constraint that the
skeletal set must span the full set and enable reconstruction of all the images and 3D points in the
70
full set using pose estimation and triangulation.
I formulate this problem by representing the joint uncertainty or covariance of a collection of
images using a weighted version of the image connectivity graph, where each edge encodes the rel-
ative uncertainty in the positions of a pair of images. The distance between two images in this graph
represents an approximate measure of the amount of uncertainty or information content connecting
the images. The problem is then to determine a sparse skeletal subgraph with as many leaf nodes
as possible (i.e., a small set of core interior nodes) that spans the full graph and preserves distances
between images as much as possible. This formulation allows us to achieve a provable bound on the
full covariance of the reduced skeletal reconstruction.
To achieve the desired bound on covariance, for every pair of images, I compare their estimated
relative uncertainty in the original graph to the uncertainty in the skeletal graph, and require that the
latter be no more than a xed constant t times the former. While this problem is NP-complete, I
develop a fast approach that guarantees this constraint on the covariance and in practice results in a
dramatic reduction in the size of the problem.
My experiments show that this approach increases efciency for large problems by more than
an order of magnitude, still reconstructs almost all of the recoverable parts of an input scene, and
results in little or no loss of accuracy in the reconstruction.
Related work My approach is closely related to research on intelligent subsampling of video
sequences for SfM to improve robustness and efciency [46, 101, 119]. The main difference in
my work is that I operate on unordered collections with more complex topology than typical linear
video sequences. Thus, my algorithm reasons about the structure of an image collection as a whole
(represented as a graph), rather than considering subsequences of an input video.
My work is also related to selecting a set of representative canonical views for an image collec-
tion for applications such as robot localization [11] or summarizing large image collections [130].
Also closely related is the work of Krause and Guestrin [80] on using the property of submodularity
to efciently select a near-optimal subset from a large set of possible observations, where the goal
is to maximize the coverage of the observations for applications such as environmental monitoring.
In contrast to these previous applications, I am optimizing over a more complex set of criteria. I
not only want to maximize coverage, but also the accuracy of the reconstruction. This leads to a
71
different set of requirements on the selected observations, described in Section 4.1.
Uncertainty analysis is of critical importance in robotics and other elds where an autonomous
agent is actively taking measurements to learn about its state and the state of its environment (called
active vision when the agent uses cameras). In mapping and exploration applications, a robot must
often keep track of not only its own state (position, orientation, etc.), but also its uncertainty in this
state, and make observations which help to minimize this uncertainty as much as possible. Thus,
researchers in active vision have studied the problem of selecting informative observations in an
online setting, in order to make most efcient use of available information. For instance, in [30],
Davison uses information value as a basis for selecting new measurements in a real-time tracking
application. My overall goal is similar, but I am given a large number of possible measurements
upfront, and select an optimal subset as a pre-process.
This chapter is organized as follows. In Section 4.1, I describe my approach for using uncertainty
as a basis for computing a skeletal graph. In Section 4.2, I discuss how I compute covariances for
image pairs, and in Section 4.3, I describe my algorithm for computing skeletal graphs. Finally, in
Section 4.4 I present results and analysis for several large image sets, and in Section 4.5 I describe
limitations and opportunities for future work.
4.1 Approximate uncertainty analysis using the image graph
Recall that SfM is the problem of building a reconstruction from a set of scene measurements (in
the form of feature correspondences) obtained from a collection of images. The algorithm described
in Chapter 3 uses all available measurements during scene reconstruction; the goal in this chapter
is to reduce the number of measurements as much as possible, while still recovering a high-quality
reconstruction.
How can we measure the quality of a reconstruction? Two desirable properties are completeness
and accuracy; a reconstruction should span all parts of the scene visible in the images and should
reect the ground-truth scene and camera positions as closely as possible.
If completeness and accuracy were the only considerations, SfM should use all available mea-
surements; the more information, the better.
1
However, efciency is also important. For Internet
1
Provided that the information ts our assumptions, e.g., independent measurements corrupted by Gaussian noise.
72
image sets, this tradeoff is particularly relevant. Such collections typically contain large numbers
of popular, and therefore redundant, views, along with some rare, but important (for reconstruc-
tion) views. If this small set of important images could be identied, the scene could potentially be
reconstructed much more quickly.
The main question is how to choose the subset of measurements to use. For simplicity, rather
than considering individual measurements, I make decisions at the level of images. For each im-
age I either include or exclude its measurements as a group. I call the set of images selected for
reconstruction the skeletal set and roughly dene the problem of selecting a skeletal set as follows:
given an unordered set of images 1 = I
1
, . . . , I
n
, nd a small subset o 1 that yields a recon-
struction with bounded loss of quality compared to the full image set. Such a reconstruction will be
an approximation to the full solution. Moreover, it is likely a good initialization for a nal bundle
adjustment, which, when run with all the measurements, will typically restore any lost quality.
To make this problem concrete, I must rst dene quality. Let us rst consider the completeness
of a reconstruction, the property that it spans the entire image collection. A complete reconstruction
contains a camera pose for each image and a 3D point for each observed point track. However, it is
sufcient to consider completeness in terms of cameras only; if all cameras are recovered, it must
be possible to recover any observed point through triangulation. A reconstruction from the skeletal
set cannot by itself be complete, as any camera not in the skeletal set will be missing. However, if a
missing camera has signicant overlap with an image in the skeletal set, it can likely be easily added
after reconstructing the skeletal set using simple pose estimation. The graph theoretic concept of
a dominating set (DS) captures this intuition. A dominating set of a graph G(V, E) is a subset of
nodes S V such that every node in V is either in S or adjacent to a node in S. An example of a
dominating set is shown in Figure 4.2.
However, it is also important that the skeletal set be connected (assuming the input set is con-
nected), so that it yields a single reconstruction rather than several disconnected reconstructions.
A connected dominating set (CDS), or a dominating set that forms a single connected component,
ts this requirement. Finally, a minimum connected dominating set is a CDS of minimum size. In
this paper, I will use an equivalent denition of a minimum CDS called a maximum leaf spanning
tree (MLST). An MLST of a graph G is the spanning tree T with the largest number of leaves; the
interior nodes of T form a minimum CDS. The concepts of a dominating set, connected dominating
73
(a) (b) (c)
Figure 4.2: Dominating sets and t-spanners. (a) An unweighted graph G. The nodes colored black
form a minimum dominating set. (b) A maximum leaf spanning tree (MLST) of G. The interior
nodes (colored black) form a minimum connected dominating set. Note that this graph is also a
4-spanner of G. (c) A 3-spanner of G. In this case, the skeletal set (the set of interior nodes, colored
black) has cardinality 5. This graph preserves distances in G more closely than the 4-spanner in (b).
set, and maximum leaf spanning tree are illustrated in Figure 4.2.
One possible way of choosing a skeletal set would thus be to nd a MLST of the image connec-
tivity graph, and take the interior nodes as the skeletal set. While this would satisfy the completeness
property, such a skeletal set would not give any guarantees about the accuracy of the reconstruction.
Thus, let us next consider accuracy, i.e., the property that the recovered cameras and points should
be as faithful to the actual scene as possible. Without ground truth, it is impossible to measure
accuracy directly. However, it is possible to estimate the uncertainty of a reconstruction, which is a
statistical estimate of the accuracy.
Uncertainty describes the looseness of a system, or how easy it is to perturb the recovered
parameters. A stiff reconstruction, where small perturbations result in a sharp increase in repro-
jection error, has relatively low uncertainty. Many reconstructions, however, suffer from drift. For
instance, when running SfM on a set of images taken along a linear path (e.g., a set of images of
building facades taken while walking down a long city street), the recovered camera path may be
slightly curved rather than perfectly straight. This occurs because information from one end of the
camera path to the other must travel a large distance, and errors in the recovered camera positions
slowly accumulate along the length of the path. Another way to put the problem is that the system
is relatively loose along a certain mode of variationbending the path slightly will not result in a
74
large increase in reprojection errorand therefore the uncertainty in the reconstruction is relatively
high. If, on the other hand, the path of images forms a closed loop, rather than a linear chain, there
will tend to be less uncertainty in the reconstruction, as information must travel at most halfway
around the loop to connect any two images.
Uncertainty in a system is often modeled with a covariance matrix for the parameters being
estimated (in this chapter, I will use the terms uncertainty and covariance interchangeably). For
SfM, the covariance is rank decient because the scene can only be reconstructed up to an unknown
similarity transform (translation, rotation, and scale). This freedom in choosing the coordinate
systemis known as the gauge freedom[150]. Covariance can only be measured in a particular gauge,
which can be xed by anchoring reference features, e.g., by xing the location and orientation of
one camera, and constraining the distance to a second camera to be of unit length. Covariance is
highly dependent on the choice of gauge. In the example of the linear chain of images, if the rst
camera is xed, there is no uncertainty in its parameters, but there may be large uncertainty in the last
camera in the chain; conversely, if the last camera were xed, the uncertainty in the rst cameras
parameters could be quite large. When xing a given camera I and computing the uncertainty in
another camera J, we are really computing the uncertainty in Js parameters relative to image I; or,
equivalently, how informative knowledge of the parameters of I is when estimating the parameters
of J.
For this reason, I do not measure uncertainty in a single gauge, but rather x each camera
in turn and estimate the resulting uncertainty in the rest of the reconstruction; I want the relative
uncertainty between all pairs of cameras to be as small as possible, to preserve the information
ow in the camera network as much as possible. Since reconstructing the scene and measuring the
actual covariance would defeat the purpose of speeding up SfM, I approximate the full covariance
by computing the covariance in reconstructions of pairs of images and encoding this information in
a graph. The global connectivity and stiffness of this graph captures information about the global
uncertainty of the SfM system. I also only consider covariance in the cameras, and not in the points,
as the number of cameras is typically much smaller than the number of points, and the accuracy of
the cameras is a good predictor for accuracy in the points.
In particular, I dene the image graph G
I
on a set of images 1 as the graph with a node for
75
C
IJ
I
J
P
Q
P
Q
(a) (b) (c) (d)
Figure 4.3: Modeling covariance with an image graph. (a) Each node represents an image, and each
edge a two-frame reconstruction. Edge (I, J) is weighted with a covariance matrix C
IJ
representing
the uncertainty in image J relative to I. (b) To estimate the relative uncertainty between two nodes
P and Q, we compute the shortest path between them by chaining up covariances (and taking the
trace at the end). In this graph, the shortest path is shown with arrows, and ellipses represent the
accumulated covariance along the chain. (c) If an edge is removed (in this case, the dashed edge),
the shortest path from P to Q becomes longer, and therefore the estimated covariance grows. (d)
A possible skeletal graph. The solid edges make up the skeletal graph, while the dotted edges
have been removed. The black (interior) nodes form the skeletal set o, and would be reconstructed
rst, while the gray (leaf) nodes would be added using pose estimation after o is reconstructed.
In computing the skeletal graph, we try to minimize the number of interior nodes, while bounding
the maximum increase in estimated uncertainty between all pairs of nodes P and Q in the original
graph.
every image and two weighted, directed edges between any pair of images with common features.
2
Without loss of generality, I assume that G
I
is connected (in practice, I operate on its largest con-
nected component). Each edge (I, J) has a matrix weight C
IJ
, where C
IJ
is the covariance in the
two-frame reconstruction (I, J) of the parameters of camera J when camera I is held xed. In
general, C
IJ
could be a full covariance matrix with entries for both position and orientation. In
my implementation, I model only the positional uncertainty of a camera, so C
IJ
is a 3 3 matrix.
Figure 4.3(a) shows an example image graph with covariance weights.
For any pair of cameras (P, Q), G
I
can be used to estimate the uncertainty in Q if P is held
xed, by chaining together covariance matrices along a path between P and Q. To compute the exact
covariance, information would need to be integrated along all paths from P to Q, and accumulated
according to the equations given by Smith and Cheeseman [134]. However, the shortest path from
P to Q gives an upper bound on the true covariance. Figure 4.3(b) illustrates this idea.
For the concept of shortest path to be well-dened, path lengths in G
I
must be scalar-valued,
rather than matrix-valued. I use the trace, tr(C), of the nal covariance matrix as the scalar length
of a path. The trace of a matrix is equal to the sum of its eigenvalues, so it expresses the magnitude
2
GI is a directed, weighted version of the image connectivity graphs introduced Chapter 3.
76
of the uncertainty. It is also a linear operator, i.e. tr(C
1
+C
2
) = tr(C
1
) +tr(C
2
), and is invariant to
rotation. Thus, adding up covariance matrices and taking the trace at the end is equivalent to adding
up the traces of the individual covariances. I can therefore convert the covariance edge weights to
scalar values, w
IJ
= tr(C
IJ
).
The goal of the skeletal sets algorithm is to sparsify G
I
as much as possible. What happens
if we start removing edges from G
I
? As we remove edges, the lengths of some shortest paths
will increase, as illustrated in Figure 4.3(c). On the other hand, removing edges from G
I
yields a
skeletal graph G
S
that is more efcient to reconstruct. I estimate this efciency by simply counting
the number of interior (i.e., non-leaf) nodes in G
S
, since once we reconstruct the interior nodes,
the leaves can easily be added in using pose estimation, and the leaves do not affect the overall
connectivity of the graph. The objective is therefore to compute a skeletal graph with as few interior
nodes as possible, but so that the length of any shortest paths (i.e., the estimated uncertainty) does
not grow by too much.
There is an inherent trade-off in this formulation: the greater the number of edges removed
from G
I
(and the greater the number of leaves created), the faster the reconstruction task, but the
more the estimated uncertainty will grow. I express this trade-off with a parameter, t, called the
stretch factor. For a given value of t, the skeletal graph problem is to nd the subgraph G
S
with
the maximum number of leaf nodes, subject to the constraint that the length of the shortest path
between any pair of cameras (P, Q) in G
S
is at most t times longer than the length of the shortest
path between P and Q in G
I
. A subgraph G
S
with this property is known as a t-spanner [5]; thus,
the problem is to nd a maximum leaf t-spanner. My algorithm for solving this problem is described
in Section 4.3. Examples of t-spanners (for an unweighted graph) are shown in Figure 4.2.
A t-spanner subsumes the property of completeness, since if a node in G
S
were to become
disconnected from the rest of the graph, some shortest path in G
S
would have innite length. Fur-
thermore, the skeletal graph will tend to preserve important topological features in G
I
, such as large
loops, as breaking such structures will dramatically increase the distance between one or more pairs
of nodes.
My approach is based on a simplied probability model. In particular, I consider only positional
uncertainty, and use shortest path covariance as a bound on the full pairwise covariance. I make
these simplications so that the cost of making decisions is signicantly smaller than the time saved
77
during reconstruction [30]. These approximations produce skeletal sets that yield remarkably good
reconstructions with dramatic reductions in computation time, as I demonstrate in the experimental
results section.
Modeling higher-level connectivity. One issue with the above problem formulation is that the
image graph G
I
is not a sufciently expressive model of image connectivity for SfM. To see why,
consider three images A, B, and C, where A and B overlap, as do B and C, but A and C do not
(violating the rule of three described in Section 3.7). These nodes form a connected subgraph
in G
I
. However, these three images do not form a consistent reconstruction, because the scale
factor between the two-frame reconstructions (A, B) and (B, C) cannot be determined (recall the
problem with the nskulla data set in Section 3.7). In order to determine the scale factor, (A, B) and
(B, C) must see at least one point in common. Therefore, any path passing through nodes A, B, C
in sequence is not a realizable chain of reconstructions; I call such a path infeasible.
To address this problem, I dene another graph, the image pair graph G
P
. G
P
has nodes for
every reconstructed pair of images, and edges between any two reconstructions that share common
features.
3
G
P
is also augmented with a node for every image, and is constructed so that a path
between two images P and Q has the same weight as the analogous path in G
I
; the only difference
is that only feasible paths can be traversed in G
P
.
More concretely, the vertex set of G
P
is (I, J) [ (I, J) is an edge in G
I
1. Since G
I
is
directed, and each connected pair of nodes (I, J) in G
I
has edges in both directions, G
P
contains
a node for (I, J) and a node for (J, I). The edge set of G
P
contains two directed edges between
each pair of reconstructions (I, J) and (J, K) that have common features. The weight on edge
[(I, J), (J, K)] is w
IJ
. The sequence (I, J), (J, K) represents a path from I to J to K; however,
only w
IJ
is represented in the edge weight. Thus, I add an extra edge from (J, K) to the special
node K representing a terminal image in a path. Edge [(J, K), K] has weight w
JK
. There are also
edges from each terminal node I to each reconstruction (I, J); these edges have weight zero. An
example of an image graph and an image pair graph showing this construction is shown in Figure
4.4. To compute a shortest path between two nodes P and Q using G
P
, I nd the shortest path
between the image nodes P and Q. Additionally, I add a constraint that such terminal image nodes
3
Note that GP is closely related to the line graph of GI [65].
78
can only appear at the ends of the path (i.e., another image node I cannot appear in the path from P
to Q), as otherwise it is possible to nd infeasible paths through G
P
.
This higher-level connectivity imposes an addition constraint on the skeletal graph: it must yield
a single, feasible reconstruction. One way to express this is to dene the embedding of a subgraph
G
S
of G
I
into G
P
as the subgraph of G
P
containing the nodes corresponding to the edges of G
S
,
and any edges between these nodes. The embedding of G
S
into G
P
must be connected for the
skeletal graph to be feasible.
Overview of the reconstruction algorithm. The basic pipeline of the skeletal sets reconstruction
algorithm is as follows:
1. Compute feature correspondence for the images, as in Section 3.1.
2. Compute a pairwise reconstruction and covariance matrices for each matching image pair,
and create the graphs G
I
and G
P
(Section 4.2).
3. Construct a feasible maximum leaf t-spanner for G
I
(Section 4.3).
4. Identify the skeletal set and reconstruct it using the reconstruction algorithm of Chapter 3.
5. Add in the remaining images using pose estimation.
6. Optionally, run a nal bundle adjustment on the full reconstruction.
4.2 Building G
I
and G
P
As in the previous chapter, the rst step is to nding correspondences by extracting SIFT features
from each image [90], matching features between each pair of images, and forming connected com-
ponents of matches to produce tracks. The matching step is extremely time-consuming on large
data sets, but researchers are making signicant progress on matching [104, 24], and I anticipate
that much faster matching techniques will soon be available.
Once correspondences have been computed for an image collection, the next step of the skeletal
sets approach is to create the image graph G
I
and the pair graph G
P
. These graphs become the
input to the skeletal graph algorithm described in the next section. The algorithm computes G
I
and G
P
in three stages: (1) it creates a two-frame reconstruction for every pair (I, J) of matching
images, removing duplicate images as it goes, (2) it computes the relative covariance matrices C
IJ
79
w
12
w
14
w
14
w
34
w
43
w
32
w
21
w
23
1
4
3
2
1,4
1,2 2,3
3,4
41
w
w
23
w
32
w
12
w
43
w
34
w
41
w
21
w
12 w
32
w
21
w
34
w
14
w
41
w
14
1,2
2,1 3,2
2,3
4,1
1,4
3,4
4,3
2
3
4
1
(a) (b) (c)
Figure 4.4: Pair graph construction. (a) An example image graph, with four images, showing the
overlaps between images (1,2), (2,3), (3,4), and (1,4). (b) A possible (simplied) pair graph for (a),
with a node for each pair of images. All pairs of reconstructions overlap, i.e., share some points in
common, except for pairs (2,3) and (3,4). (c) An augmented pair graph with edge weights shown;
the bold edges correspond to the edges in graph (b). This graph is augmented with a node for each
image, and allows for computation of lengths of feasible paths between images, with the constraint
that an image node (shown as a circle) can only appear at the ends of a pathonly the bold edges of
this graph can be used in a path, except at the very beginning and very end of the path. For instance,
the only allowable path from image 2 to image 4 is a path through nodes 2,1 and 1,4. A path that
contains image node 1 or 3 is disallowed.
and C
JI
for each pair, and (3) it checks which pairs of two-frame reconstructions overlap (and are
therefore edges in G
P
). These steps are now described in detail.
Two-frame reconstruction and duplicate detection. I rst compute a reconstruction for each
matching image pair using the ve-point relative pose algorithmof Nist er [102] inside of a RANSAC
loop. The ve-point algorithm requires both cameras to be calibrated. Therefore, I only consider
images that have a focal length estimate encoded in their Exif tags (and assume that each camera
has unit aspect ratio, zero skew, and a principal point at the image center, as before). As described
in Section 3.2.2 of the previous chapter, the focal length estimates in Exif tags are typically off by
several percent (and are occasionally completely wrong); however, I have found that they are usually
close enough to give reasonable pairwise reconstructions. After estimating the relative pose for a
80
pair, I triangulate matches that are inliers to the recovered two-frame reconstruction and run bundle
adjustment using the Sparse Bundle Adjustment package (SBA) [89]. If the average reprojection
error of a reconstruction is too high (I use a threshold of 0.6 pixels), I discard the reconstruction,
as these have usually been misestimated (due, for instance, to an erroneous focal length, or to other
factors described in Section 3.7, Bad initialization).
Once a pair has been reconstructed, I check whether the images are near-duplicates, i.e., whether
they have very similar image content, or if one is subsumed by other. The reason for this check is that
many images in Internet collections are very similar (and occasionally exactly identical), and much
work can be saved by identifying such near-duplicates early on. I consider image J to duplicate
image I if (a) they represent views taken very close together,
4
and (b) all the images connected to
J in G
I
are also connected to I (so that node J can be collapsed into node I without affecting the
connectivity of the graph). If both of these criteria hold, it is likely that nearly all of the geometric
information in image J is also present in image I, and the algorithm removes J from consideration
(it will be added back in at the very end of the reconstruction process).
Without duplicate detection, the total number of pairs processed would be equal to the number
of matching images, which could be as high as O(n
2
) for n images. With duplicate detection, the
algorithm can often avoid processing many pairs. For instance, in the extreme case where all n
images are the same, I will only process n pairs, rather than n
2
. In practice, the total number of
pairs processed depends on the order in which they are considered; if many duplicates are removed
early on, fewer pairs will be processed. Therefore, observing that images that are more similar also
tend to have more matching features, I sort the pairs by number of matches, and consider those with
the most matches rst. For the Internet collections I tested, typically about a third of the images are
removed as duplicates.
Covariance estimation. Once I have reconstructed a (non-duplicate) pair (I, J), I estimate the
relative covariances C
IJ
and C
JI
of the two camera positions. During bundle adjustment, SBA
uses the Schur complement to factor out the 3D point parameters and compute the Hessian H
C
of the reduced camera system [150]. To estimate C
IJ
and C
JI
, I invert H
C
and select the blocks
4
In particular, if the distance, dc, between the camera centers is small compared to the median distance, dp, between
the cameras and the reconstructed points (I use dc < 0.025dp).
81
corresponding to the camera positions of I and J. However, because of the gauge freedom, H
C
is
a singular matrix, so I rst add constraints to x the gauge and make H
C
invertible. In particular,
to estimate C
IJ
, I x the position and orientation of the reconstruction by constraining camera I
to be at the origin with an identity rotation. I also x the scale of the reconstruction by adding a
soft constraint on the mean of the 3D points encouraging the mean point to stay at its current depth
with respect to the rst camera (the mean is computed after removing very distant points; in my
implementation, I remove all points beyond 1.2 times the 90
th
percentile distance). An additional
penalty term equal to the difference between the target depth and current depth (weighted by a
constant; I use 2) is added to the objective function to enforce this constraint.
5
With these constraints, H
C
is invertible, and can be used to compute C
IJ
. C
JI
is computed
analogously by xing camera J (in general, the two covariances are not identical, i.e., C
IJ
,= C
JI
).
Constructing the pair graph G
P
. After computing covariances for every image pair, I construct
the pair graph G
P
. Recall that every node of G
P
represents a pairwise reconstruction, and that an
edge connects every pair of overlapping reconstructions (I, J) and (J, K). The main remaining
task is to determine this set of edges, i.e., to compute which pairs of reconstructions are connected
by common points. To do so, I consider each triple of images (I, J, K) where (I, J) and (J, K)
are reconstructions. I nd the intersection of the point sets of (I, J) and (J, K), then use absolute
orientation [69], inside a RANSAC loop, to nd a similarity transform T between the two pairwise
reconstructions. If there are at least a minimum number of inliers to T (I use 16), the two directed
edges ((I, J), (J, K)), and ((K, J), (J, I)) are added to G
P
. The scale factors s
ijk
and s
kji
(where
s
ijk
s
kji
= 1) between the two reconstructions are also stored with each edge, so that the algorithm
can properly align the scales of adjacent reconstructions when computing shortest paths in G
P
, as
described in Section 4.3.4.
Finally, I augment the graph G
P
with nodes and edges for each image, as described in Section
4.1. Now, with G
I
and G
P
in hand, we are ready for the skeletal set algorithm.
5
Another approach to xing the scale is to constrain the distance between the cameras to be of unit length; however,
this will remove any uncertainty in the translation direction, and can therefore bias the covariance. In particular, for
image pairs where the translation is primarily forward, i.e. one camera is roughly in front of another, the uncertainty
is typically largest in the translation direction, therefore zeroing out this uncertainty will make such pairs look more
certain than they really are.
82
4.3 Computing the skeletal set
As mentioned in Section 4.1, I formulate the problem of computing a skeletal set as that of nding
a (feasible) maximum leaf t-spanner of G
I
, called the skeletal graph, G
S
. The feasibility criterion
requires that the embedding of G
S
in G
P
must be connected. To ensure that this constraint is satis-
ed, my algorithm maintains data structures for both G
S
and its embedding in G
P
when computing
G
S
. Once G
S
is found, the skeletal set o is found as the set of interior nodes of G
S
.
Unfortunately, the problemof computing a minimumt-spanner for general graphs is NP-complete
[18], so it is unlikely that an exact solution to the maximum leaf t-spanner problem can be found
efciently. To get around this, I have developed an approximation algorithm for computing G
S
which consists of two steps. First, a spanning tree T
S
of G
I
is constructed. The construction of
T
S
balances computing a tree with a large number of leaves (a maximum leaf spanning tree) with
computing a tree with a small stretch factor (a t-spanner). Because no tree may have the desired
stretch factor of t, the second step is to add additional edges to T
S
to satisfy the t-spanner property.
I now describe these two steps in detail.
4.3.1 Constructing the spanning tree
I begin by describing a simple, greedy approximation algorithm for computing a maximum leaf
spanning tree (MLST) proposed by Guha and Khuller [61]. The idea behind the algorithm is to
grow a tree T
S
one vertex at a time, starting with the vertex of maximum degree.
Basic MLST algorithm. The algorithm maintains a color for each node. Initially, all nodes are
unmarked (white), and the algorithms proceeds as follows:
1. Let T
S
be initialized to the empty graph. Select the node v of maximum degree. Add v to T
S
,
and color v black.
2. Add every unmarked neighbor of v, and the edge connecting it to v, to T
S
and color these
neighbors gray.
3. Select the gray node v with the most unmarked neighbors, color it black, and go to step 2,
until all nodes are black or gray.
83
I rst modify this algorithm to ensure that the constructed tree is feasible. To do so, I maintain a
parent for each node (except the rst). The parent P(v) of a node v is the node that caused v to be
colored gray. In step 2 of the algorithm, I only color a neighbor u of v gray if the path (P(v), v, u)
is feasible. Similarly, in step 3, when counting unmarked neighbors of a node v I only consider
those for which (P(v), v, u) is feasible. These constraints guarantee that the reconstructed tree is a
feasible reconstruction.
4.3.2 Considering edge weights.
The basic MLST algorithm nds a spanning tree with a large number of leaves, but the stretch
factor of this tree could be arbitrarily high, and many extra edges may need to be added to reduce
the stretch factor to t. Thus, the spanning tree algorithm should ideally produce a tree with a stretch
factor as close to the target value of t as possible, so as to minimize the number of extra edges
required. It should therefore select nodes and edges that not only have high degree, but which are
also critical for keeping distances between nodes as short as possible. Some images and edges are
naturally more important than others. For instance, some nodes and edges might be on a large
number of shortest paths, and removing them may result in large growth in distances between many
pairs of nodes. On the other hand, some edges in a graph may not be along any shortest path, and
are therefore relatively unimportant. Therefore, I integrate a notion of node and edge importance
into the algorithm.
There are many possible ways of measuring the importance of a node to the global connectivity
of a graph. Many measures of the centrality of a node have been proposed, including betweenness
centrality, which measures the number of shortest paths that pass through a node, eigenvector cen-
trality (a variant of which is used in Googles PageRank algorithm [106]), and closeness, which
measures the average distance from a node to every other node. These measures can be expensive
to compute, so in my algorithm, I take a very simple, local approach. I rst measure the importance
of an edge (I, J) by computing the length of the shortest feasible path between I and J in G
I
(I
denote this length d
f
(I, J; G
I
)), and dividing by the weight of (I, J) itself:
imp(I, J) =
d
f
(I, J; G
I
)
w
IJ
.
84
0.8
0.1
0.2
0.1
I J
0.5
1.0
1.0
1.0
K L
Figure 4.5: Edge importance. A four-node region of an undirected graph is shown. The edge
weights are shown in normal typeface, and the importance of an edge is shown in red italics. Three
of the edges ((I, K), (J, K), and (K, L)) have an importance of 1, because each of these edges is
the shortest path between its endpoints. On the other hand, the shortest path between I and J is the
sequence (I, K, L, J), which has length 0.4, while the length of edge (I, J) itself is 0.8. Because
there is a path 0.5 times the length of (I, J), imp(I, J) = 0.5. Edge (I, J) would be removed from
the graph if the importance threshold > 0.5.
If the edge (I, J) is itself a shortest path between I and J, imp(I, J) = 1. Otherwise, imp(I, J) <
1, and the longer the edge is compared to the shortest path, the smaller imp(I, J) will be. Some
pairwise reconstructions (I, J) are naturally ill-conditioned, and a much higher certainty can be
achieved via a detour through one or more other images. Such edges receive a low importance
score. This denition of importance is illustrated in Figure 4.5.
While this is a much more local denition of importance than the others mentioned above, it has
several advantages. First, it is relatively efcient to compute. Second, it gracefully handles edges
that are almost shortest paths (unlike, e.g., the betweenness centrality measure, which would assign
an importance of zero to an edge that is almost, but not quite, a shortest path). Finally, there is a
connection between this measure of importance and the stretch factor of a graph, described below
(Setting the importance threshold).
Before running the basic MLST algorithm, I remove edges from G
I
that have an importance
score lower than a threshold , forming a pruned graph G
I

. The degree of a node in G

I

is then the
number of incident important edges, and is a better predictor for how important the node is than
its raw degree in G
I
.
Setting the importance threshold. How should the importance threshold be set? The tradeoff
is that with a very small threshold, very few edges will be pruned and the MLST algorithm will
85
try to maximize the number of leaves in T
S
, without considering the stretch factor, as before. With
a larger threshold, more edges will be pruned, and it may be more difcult to create a tree with a
large number of leaves, but the stretch factor of the tree will likely be smaller. Is there a connection
between t and ?
Yes. Consider the case when t = . Any spanning tree of G
I
will satisfy this stretch factor, so
we need not prune any edges ( = 0). Now consider a nite value of t, and suppose an edge (I, J)
has an importance score imp(I, J) <
1
t
. I claim that this edge will not be included in a minimum
t-spanner of G
I
. To see why, rst notice that, by the denition of importance, there must be a path
between I and J in G
I
that has length <
w
IJ
t
. Thus, a t-spanner G
S
of G
I
must contain a path
of length at most t between I and J. Edge (I, J) does not satisfy this requirement, because
w(I, J) > t . Therefore there must exist another, shorter path between I and J in G
S
. As a
result, no shortest path in G
S
will contain edge (I, J), because a shorter path can be obtained with
a detour through . Any t-spanner containing edge (I, J) will therefore not be minimal, as (I, J)
can be removed without changing the stretch factor.
Given this, we can safely remove from G
I
all edges (I, J) such that imp(I, J) <
1
t
before
constructing the spanning tree T
S
; none of these edges can appear in G
S
. In my implementation, I
use a larger importance threshold of =
4
t
(clamped to a maximum of = 1), thereby removing a
larger number of weak edges.
4.3.3 From MLST to t-spanner
The tree T
S
computed above spans the entire graph, but may not be a t-spanner. To guarantee that
the stretch factor of the skeletal graph is at most t, additional edges may need to be added, forming
a graph G
S
with cycles. In order to determine which, if any, edges to add, the algorithm must rst
test whether the target stretch factor has been met. How can we verify that a graph is a t-spanner?
One way would be to take every pair of nodes I and J, compute the shortest paths between I and J
in both G
I
and G
S
, and compare the lengths of the two shortest paths. However, it is not necessary
to check shortest paths between all pairs of nodes to verify that a subgraph is a t-spanner; it sufces
to only check shortest paths between all neighboring nodes (I, J) in G
I
. The reasoning behind this
is that if the distance between all neighboring nodes in a graph is dilated by at most a constant factor
86
t, the distance between any two nodes must also be dilated by at most a factor t, because each edge
on the original shortest path can be replaced by a new path at most t times longer.
I therefore enumerate all edges of G
I
not included in the tree T
S
. For each edge (I, J), I
compute the feasible distance between I and J in both graphs, denoting these distances d
f
(I, J; G
I
)
and d
f
(I, J; G
S
). If t d
f
(I, J; G
I
) < d
f
(I, J; G
S
), I add edge (I, J) to G
S
; otherwise, I omit
the edge. Once this procedure is complete, all neighboring nodes (I, J) of G
I
must satisfy the
t-spanner property, and therefore G
S
must be a t-spanner of G
I
.
The set of edges added to G
S
depends on the order in which the edges are processed, since
adding a single edge can affect many shortest paths, obviating the need to add other edges later on.
Therefore, I rst consider edges between nodes that are already on the interior of G
S
(i.e., the black
nodes of the MLST algorithm). I then follow the heuristic of

Althofer, et al., [5] and consider the
remaining edges in order of increasing covariance weight.
Putting it all together. Once G
S
has been augmented with the necessary edges, the skeletal set
o is selected as the set of non-leaf nodes of G
S
. The skeletal set is then reconstructed with the
incremental structure from motion technique described in the previous chapter. Next, the remaining
(leaf) images are added using pose estimation [68], forming a full reconstruction. Bundle adjustment
is then optionally run on the full reconstruction.
In summary, the skeletal sets algorithm has the following steps:
1. Compute feature correspondences for the images.
2. Compute a reconstruction and covariances for each matching image pair, and remove dupli-
cate images (Section 4.2). Create the graphs G
I
and G
P
.
3. Compute an importance score for each edge of G
I
, and prune edges with low importance
(Section 4.3.2)
4. Construct a MLST from G
I
(Section 4.3.1).
5. Add edges to guarantee the stretch factor (Section 4.3.3).
6. Identify and reconstruct the skeletal set.
7. Add in the remaining images using pose estimation.
8. Optionally, run bundle adjustment on the full reconstruction.
87
4.3.4 Implementation of shortest path computation in G
P
.
Shortest paths are computed in two steps of the skeletal set algorithm: (1) step 3, computing the
importance of each edge in the graph, and (2) step 5, determining which edges must be added to T
S
to achieve the stretch factor. These paths are computed in the graph G
P
so that only feasible paths
are considered.
The shortest path between two nodes in a graph G = (V, E) can be computed using Dijkstras
algorithm [35] in O([E[ + [V [ log [V [) amortized time by using a Fibonacci heap [48] to nd the
minimum-distance node in each iteration of the algorithm. I compute shortest paths in G
P
with a
modied version of this algorithm. The main difference comes from the fact that the covariance
weights on the edges of G
P
are derived from reconstructions in different coordinate systems, so the
covariances are not directly comparable. Thus, I use the scale factors computed when constructing
the pair graph to scale the edge weights during the shortest path computation (the edge weights
are scaled by the square of the scale factors, as the trace of a covariance matrix grows with the
square of the scene scale). In addition, for each image I, care must be taken to make sure that all
outgoing edges have weights measured in the same coordinate system. Therefore, as a preprocess
to the skeletal sets algorithm, for each image I I select a reconstruction (I, J) to be the canonical
coordinate system for I, and align all other reconstructions (I, K) to (I, J). Not all reconstructions
(I, K) may overlap with (I, J), but through transitivity most can be aligned (I remove any remaining
unaligned reconstructions).
Note that the shortest path algorithmcan often be terminated early (i.e., it is not always necessary
to nd the exact shortest path between two nodes). In the importance pruning stage, if at any time
a path between I and J in G
I
shorter than w
IJ
is found, (I, J) can immediately be pruned.
Similarly, in the stretch factor stage, if any path between I and J in T
S
shorter than t w
IJ
is found,
we know that (I, J) can be omitted from T
S
.
4.4 Results and discussion
I have evaluated my algorithm on several large Internet photo collections of famous world sites (St.
Peters Basilica, Stonehenge, the Pantheon, the Pisa Duomo, Trafalgar Square, and the Statue of Lib-
erty). I obtained these data sets by doing keyword searches on Flickr and downloading the results.
88
(a) (b) (c) (d)
(e)
Figure 4.6: Reconstruction of Stonehenge. (a) The full image graph for Stonehenge and (b) the
skeletal graph (for t = 16). The black (interior) nodes of (b) comprise the skeletal set, and the
gray (leaf) nodes are added in later. Note that the large loop in the graph is preserved. (The graph
layouts are not based on the physical position of the cameras, but are created with the neato tool in
the Graphviz package). (c) Aerial photo. (d-e) Two views of the Stonehenge reconstruction. The
reconstructions show the recovered cameras, rendered as small frusta, in addition to the point cloud.
(a) (b) (c) (d)
Figure 4.7: Reconstruction of the interior of St. Peters. (a) The full image graph for St. Peters
and (b) our skeletal graph (for t = 16). In this graph, white nodes represent images found to be
duplicates. These nodes are removed before computing the skeletal graph. (c) Overhead photo of
St. Peters. (d) The nal reconstruction.
89
Table 4.1: Data sets and running times. Each row lists: name, the name of the scene; # images, the
number of input images; largest cc, the size of the largest connected component of the image graph;
[ o [, the size of the computed skeletal set; #reg full, the number of images registered in the full
reconstruction; #reg o, the number of images registered in the skeletal set reconstruction; rt full,
the running time of the full reconstruction; rt o, the running time of the skeletal set reconstruction,
including computing the pairwise reconstructions and the skeletal graph; rt o+BA, the running time
of the skeletal set reconstruction plus a nal bundle adjustment. The columns in bold represent
the running times for the full reconstruction with both the baseline method and the skeletal sets
algorithm.
Name # images largest cc [ o [ #reg full #reg o rt full rt o rt o+BA
Temple 312 312 54 312 312 46 min 58.5 min 63.5 min
Stonehenge 614 490 72 408 403 276 min 14 min 26 min
St. Peters 927 471 59 390 370 11.6 hrs 3.54 hrs 4.85 hrs
Pantheon 1123 784 101 598 579 108.4 hrs 7.58 hrs 11.58 hrs
Pisa1 2616 1671 298 1186 1130 17.8 days 14.68 hrs 22.21 hrs
Pisa2 1112 1110 352 1101 1093 18.5 days 32.9 hrs 37.4 hrs
Trafalgar 8000 3892 277 - 2973 > 50 days 17.78 hrs 30.12 hrs
Liberty 20931 13602 322 - 7834 > 50 days 15.59 days N/A
a
a
Running bundle adjustment on the full reconstruction was not possible due to memory constraints.
I also tested on a second collection of images of the Pisa Duomo taken by a single photographer
with the express purpose of scene reconstruction (I refer to the Internet collection as Pisa1, and the
single-photographer collection as Pisa2). Finally, I evaluated the algorithm on the Temple data set.
This 312-image collection is taken from the multi-view stereo evaluation data of Seitz, et al. [128],
and was captured in the lab using the Stanford spherical gantry, so the camera poses are known (the
camera used was also calibrated, so the intrinsics are known as well). The images are regularly
sampled on a hemisphere surrounding a plaster reproduction of the Temple of the Dioskouroi.
I reconstructed each data set using the skeletal graph algorithm with a stretch factor t = 16.
Visualizations of the full and skeletal image graphs for the Stonehenge data set are shown in Fig-
ure 4.6, for St. Peters in Figure 4.7 and for the Pantheon in Figure 4.8. Note that the skeletal graphs
are much sparser than the original graphs, yet preserve the overall topology. For instance, the large
loop in the full image graph for Stonehenge is retained in the skeletal graph.
Figure 4.8 shows overhead views of the Pantheon during several stages of our algorithm; note
that both the inside and outside are reconstructed (connected by images that see through the door-
90
(a) (b) (c) (d) (e)
Figure 4.8: Reconstructions of the Pantheon. (a) The full image graph for the Pantheon and (b) our
skeletal graph. The black (interior) nodes of (b) comprise the skeletal set, and the gray (leaf) nodes
are added in later. The Pantheon consists of two dense sets of views (corresponding to the inside
and outside), with a thin connection between them (views taken outside that see through the door).
Note how the skeletal set preserves this important connection, but sparsies the dense parts of the
graph. (c) Reconstruction from the skeletal set only. (d) After using pose estimation to register the
remaining images. (e) After running bundle adjustment on (d).
way). Views of the reconstruction of Stonehenge are shown in Figure 4.6, St. Peters in Figure 4.7,
Pisa1 in 4.9, Pisa2 in Figure 4.10, and Trafalgar Square in Figure 4.11.
Table 4.1 summarizes the running time for each data set (excluding the correspondence esti-
mation stage). The running times reported for my algorithm are for the entire skeletal sets pipeline
(computing pairwise reconstructions, building the skeletal graph, and reconstructing the scene). For
Trafalgar and the Statue of Liberty (the largest sets), the baseline method was still running after 50
days.
The results show that the skeletal sets method (with the exception of the Temple dataset, to be
discussed shortly) reconstructs scenes signicantly faster than the baseline method, and the perfor-
mance gain generally increases dramatically with the size of the data set. The speedup ranges from
a factor of 2 for St. Peters, to a factor of at least 40 for Trafalgar Square. At the same time, the
algorithm recovers most of the images reconstructed by the baseline method. A few images are lost;
most of these are very tenuously connected to the rest, and can mistakenly be pruned as infeasible
while building the skeletal graph. My method also works well on the set taken by a single person
(Pisa2), although the fraction of images in the skeletal set for this collection is higher than for the
91
Figure 4.9: Several views of the Pisa1 reconstruction.
Figure 4.10: Overhead photo and view of the Pisa2 reconstruction.
92
Figure 4.11: Views of the Trafalgar Square reconstruction, with images for comparison.
93
(a)
(b)
Figure 4.12: Views of the Statue of Liberty reconstruction. (a) An overhead view of the reconstruc-
tion centered on the Statue. The point cloud itself is in the very center of the image, and the small
black triangles around the point cloud represent recovered cameras. Many photos were taken either
on Liberty Island itself or from boats out in the water. (b) Overhead view of the full reconstruction,
including parts of Manhattan on the right.
94
0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0 10 20 30 40 50 60 70 80
F
r
a
c
t
i
o
n

o
f

i
m
a
g
e
s
i
n

s
k
e
l
e
t
a
l

s
e
t
Stretch factor
0
0.5
1
1.5
2
2.5
0 10 20 30 40 50 60 70 80
E
r
r
o
r

(
m
)
Stretch factor
Med. error (before BA)
Med. error (after BA)
Figure 4.13: Stretch factor analysis for St. Peters. Left: stretch factor vs. number of nodes in
the skeletal set. Right: median error (in meters) in camera position for reconstructions before and
after nal bundle adjustment. As the stretch factor increases, the error before bundling increases,
but applying bundle adjustment results in a low error level (around 6-8cm; note that the nave of the
cathedral is about 24m across), even for stretch factors as large as 30.
Internet sets, due to the more regular distribution of photos.
For most of the data sets, the algorithm spent more time in reconstruction than in building the
skeletal graph; for a few particularly dense sets (e.g., the Pantheon), the preprocessing took more
time (for the Statue dataset, the pairwise reconstructions took over 90% of the reconstruction time).
This was also true for the Temple data set, the only set for which the baseline method outper-
formed the skeletal set algorithm. Because of the high density of images, there are a large number
of overlapping pairs, yet few pairs that match my denition of duplicate. Hence, a large amount
of time (approximately 55 minutes) was spent creating pairwise reconstructions. In contrast, once
the skeletal set was computed, the full model was reconstructed in just 9 minutes, compared to 64
minutes for the baseline reconstruction. The baseline reconstruction still took less time, however,
due to the length of time taken in computing pairwise reconstructions. There is signicant opportu-
nity to speed up this step, for instance, through parallelization (each pairwise reconstruction can be
computed independently), or through a more efcient implementation of RANSAC.
Next, I analyze the tradeoff between the stretch factor t and the accuracy of the reconstruction.
To evaluate this tradeoff, I took one of the data sets (St. Peters) and reconstructed it with multiple
values of t. There is no ground truth reconstruction for this data set, so I used the reconstruction
obtained from running the baseline method on the full image set as a standard by to measure ac-
95
0
0.2
0.4
0.6
0.8
1
0 10 20 30 40 50 60
F
r
a
c
t
i
o
n

o
f

i
m
a
g
e
s
i
n

s
k
e
l
e
t
a
l

s
e
t
Stretch factor
0
0.1
0.2
0.3
0.4
0 10 20 30 40 50 60
E
r
r
o
r

(
c
m
)
Stretch factor
Avg. error (before BA)
Avg. error (after BA)
Avg. error (baseline)
Figure 4.14: Stretch factor analysis for the Temple data set. The graph on the left plots the number
of images in the skeletal set for several values of the stretch factor t. The graph on the right plots
the average distance between the reconstructed and ground truth camera centers for these values
of t, showing average error (in cm) both before and after a nal bundle adjustment. Note that the
error before bundle adjustment, while noisy, tends to increase with the stretch factor, but the error
after bundle adjustment stays roughly constant. I also ran SfM with the full set of input images,
and plot the error as a baseline. This data set is very controlled and regularly sampled compared
to the Internet collections, and the behavior as t changes is somewhat different than the St. Peters
collection. For instance, even for the largest values of t we tried, the initialization provided by the
skeletal reconstruction was close enough to the correct solution for the nal bundle to pull it to the
right place.
curacy. (This baseline reconstruction, which uses all available information, should in theory have
the highest accuracy.) For each value of t, I rst aligned the resulting reconstruction to the baseline
reconstruction by nding a similarity transform between corresponding points. I then computed the
distances between corresponding cameras, both before and after a nal bundle adjustment. Figure
4.13 shows the results, plotting the size of the skeletal set, and the median error in camera position,
for several values of t. As t increases, the size of the skeletal set decreases, and the error before
the nal bundle adjustment increases. However, applying the nal bundle results in a low, relatively
constant error level (in this case, a median error between 6-8cm for a building about 24m in width),
even for stretch factors as large as 30, at which point only 10% of the images are used in the skeletal
set. For even larger stretch factors, however, the bundled solution begins to degrade, because the
initialization from the skeletal set is no longer good enough to converge to the correct solution.
I then ran the same experiment on the Temple dataset (this time comparing to the known ground
truth, rather than the results of the baseline method). Figure 4.14 shows the results. The error before
96
the nal bundle adjustment is noisy, but tends to generally increase with the stretch factor; the error
after bundle adjustment stays roughly constant, even for the largest stretch factors I tried. This data
set is very controlled and regularly sampled compared to the Internet collections, and the behavior
as t changes is somewhat different than the St. Peters collection. For instance, even for the largest
values of t I tried, the initialization provided by the skeletal reconstruction was close enough to the
correct solution for the nal bundle to pull it to the right place.
4.5 Conclusion
I have presented an algorithm for reconstructing Internet photo collections by computing a skeletal
graph, and shown that this method can improve efciency by up to an order of magnitude, with
little or no loss in accuracy. There are also many interesting avenues for future work. First, my
experiments indicate that a bottleneck in the algorithm is the computation of the two-image recon-
structions and covariances for each pair of matching images. This step could easily be made much
faster through paralellization, but are also other potential ways to improve efciency. For instance,
if a lower bound on, or an approximation to, the covariance for a pair of images could be computed
quickly, it might be possible to compute the skeletal set and the pairwise reconstructions at the same
time, only computing exact edge weights when required. An online approach to creating the skeletal
set would also be interesting in cases where new images are coming in all the time (as is the case
with the Internet), and should be incorporated into existing reconstructions. Many of these new
images would likely contain redundant information, and need not be added to the skeletal set, but
others may be desirable to strengthen or extend the current skeletal set.
It would also be interesting to create a more sophisticated model of uncertainty, e.g, taking
uncertainty in camera orientation, and perhaps in scene structure, into account, or by considering
multiple paths between image pairs. It could also be fruitful to work with triples, rather than pairs,
for convenience in representing connectivity and improved robustness.
97
Part II
NAVIGATION AND VISUALIZATION OF 3D PHOTO COLLECTIONS
In Part I of this thesis, I described my approach to recovering the geometry of large, unstructured
photo collections. This capability can enable a host of applications, e.g., taking 3D measurements
of objects, studying the behavior of photographers, and analyzing images for use in forensics. How-
ever, one of the most exciting applications, and the focus of Part II, is to automatically recreate the
experience of the worlds famous places from photos on the Internet and to build new 3D interfaces
for immersively exploring and interacting with our world.
What should such interfaces look like? The answer is not immediately clear, because these
reconstructions really combine two intertwined artifactsphotos and sceneseach of which are
interesting to explore in isolation. In a photo collection, each individual photo was carefully com-
posed and captured to tell a story or record a moment of some sentimental or aesthetic importance
to the photographer. The photos themselves are therefore worthy of examination, not only to the
photographer, but potentially to their larger circle of friends and family or to the world as a whole.
Photo-sharing sites like Flickr and SmugMug have built up communities of people who share pho-
tos with one another and view, comment on, and tag each others photos, validating the idea that
photographs can be artifacts of value to many people, even complete strangers. The second artifact
is the underlying scene that is reconstructed. Scenesexamples of which include the Taj Mahal,
the Pantheon in Rome, Yosemite National Park, and ones own homeare also interesting and
can be meaningful for aesthetic, historical, and personal reasons. Thus, within these reconstructed
collections, there are two interrelated tasks to consider:
1. Photo browsing: getting an overview of a photo collection, nding specic images, and
viewing and sharing progressions of photos (telling stories).
2. Scene exploration and understanding: building a mental map of a place, nding specic
objects, rooms, or views, and learning the names of and facts about particular parts of a scene.
In my work, I have created interfaces that address both of these tasks together. For each of these
98
tasks, I argue that there are elements of the other that can be benecial. For photo browsing, placing
the user inside of a 3D scene to immersively interact with a photo collection can aid in understand-
ing the physical relationships between different photos and allow for the creation of new types of
geometric search controls that help nd photos of specic objects or viewpoints. I also believe that
putting photos together in a single, consistent 3D world can help strengthen photo communities like
Flickr, by placing and relating the communitys photos together in a way that makes it easier see
how they t together. On the other hand, for the task of scene exploration, the very fact that photos
are meant to be interesting means that the statistics of large photo collections, taken by many dif-
ferent people, can reveal a great deal of information about the underlying scene; for instance, what
parts of the scene seem important, how people tend to move around in the scene, and the locations
of individual objects.
I explore these issues in the next two chapters, which describe two new systems for exploring 3D
photo collections. In Chapter 5, I describe the Photo Tourism system, an immersive photo browser
that uses 3D information to provide new geometric controls for exploring and annotating large
photo collections. In Chapter 6, I describe Pathnder, a scene exploration tool that uses large photo
collections for rendering, and which automatically discovers interesting, scene-specic navigation
controls by analyzing the distribution of views in the input collections. While these two systems
look at two sides of a coin, they are very much relatedthey both take the same input, and neither
one is purely about photo browsing, or purely about scene understanding. Rather, both examine the
ways in which these two tasks complement each other.
Two critical components of both systems are rendering (how to effectively display the scene
and create good transitions between images) and navigation (how to browse the photos and ex-
plore the scene). For the problem of rendering, the sparse points produced by SfM methods are by
themselves very limited and not always effective for rendering a scene. Nonetheless, I demonstrate
that camera information and sparse geometry, along with new morphing, image stabilization, and
non-photorealistic proxy rendering techniques, are sufcient to provide compelling scene render-
ings and view interpolations. Likewise, the unstructured nature of the photo collections can make it
difcult to provide the free-viewpoint navigation controls found in games and 3D viewers, or even
the simplied 3D controls in applications like Google Street View. For both systems, I describe new
techniques for creating simple, easily-understood controls for unstructured collections.
99
Chapter 5
PHOTO TOURISM: AN IMMERSIVE 3D PHOTO BROWSER
How can we build a better photo browser for collections of photographs for which we know
exactly where each photo was taken, and in which direction it was looking, and have a sparse 3D
model of the scene? The reconstruction algorithms of Part I of this thesis present us with this
opportunity, and this chapter presents a 3D photo browser called Photo Tourism that addresses this
question.
In many ways Photo Tourism is primarily a photo browser. It has tools for searching for images,
playing slideshows, and tagging images, and the emphasis of the system on moving the user from
one photo to another. However, it combines these common photo browsing features with several
elements of 3D navigation. First, the primary view is a window into a virtual world where the
user is moving around in 3D among the reconstructed photos. Second, the search controls include
new geometric and object-based controls, which make it easier to nd photos of particular views
or objects. Finally, the reconstructed 3D information is used to produce attractive 3D transitions
between images, which highlight the spatial relationships between photos. Such transitions can be
much more powerful and evocative than the cross-dissolves used in traditional photo browsers.
In addition to navigation and rendering, semantics plays an important role in photo browsing. It
is often useful or interesting to learn more about the content of an image, e.g., which statue is this?
or when was this building constructed?things someone familiar with the content of the photos
might be able to tell us. A great deal of annotated imagery of this form already exists in guidebooks,
maps, and Internet resources such as Wikipedia [158] and Flickr. However, the images in your own
personal collection, a random photo on the web, or the image you may be viewing at any particular
time (e.g., on your cell phone camera) may not have such annotations, and a tour guide may not
be readily available. A very powerful feature of Photo Tourism is the ability to transfer annotations
automatically between images, so that information about an object in one image can be propagated
to all other images that contain that same object.
100
System overview. The input to Photo Tourism is the scene and camera geometry recovered from
the structure from motion (SfM) algorithms described in Chapters 3 and 4. Recall from Chapter 3
that this geometry consists of:
1. A set of cameras, C = C
1
, C
2
, . . . , C
k
, each with a corresponding image image I
j
and a
set of viewing parameters C
j
.
2. A set of 3D points P = p
1
, p
2
, . . . , p
n
, each with a color and a position X
i
.
3. A set of line segments L = l
1
, l
2
, . . . , l
m
.
4. A 3D plane, Plane(C
i
), for each camera C
i
, t to the 3D points seen by that camera.
5. A 3D plane, CommonPlane(C
i
, C
j
), for each pair of overlapping cameras C
i
and C
j
.
Photo Tourism consists of three main components. The rst is the rendering engine. While the
user is often looking at a single photo, as in any other photo browsing tool, the user really exists in
a 3D world and can view the scene from any angle. In addition, when the system transitions from
one image to another, it physically moves users between the two, as if they were ying between the
images in 3D. In addition to camera motion, these transitions use a specialized rendering algorithm
for morphing between pairs of photographs. Users are not strictly conned to the photos, however,
and can also choose to move freely through the scene, which is rendered using the recovered points
and line segments. These rendering techniques are described in Section 5.1.
The second component of Photo Tourism consists of the navigation interface for exploring and
searching for photos. There are four kinds of controls: zooming controls for nding details, similar
views, or broader perspectives of a given photo; directional controls for nding photos that show
more in a particular direction; object-based controls for nding views of a specic object; and regis-
tered slideshows for viewing a sequence of related images. These navigation controls are described
in Section 5.2.
The nal component of Photo Tourism consists of tools for scene annotation and enhancement.
A user can label an object in one image and have that label propagate to all other relevant images. In
addition, a user can also augment a scene with additional photos, which are automatically registered
to the reconstructed scene. These capabilities are described in Section 5.3. Finally, Section 5.4
(and especially the companion video on the project webpage [108]) presents results, and Section 5.5
concludes with a discussion of the system.
101
(a) (b)
Figure 5.1: Screenshots from the Photo Tourism viewer. Left: when the user visits a photo, that
photo appears at full-resolution, and information about it appears in a pane on the left. Results of
search operations appear in the thumbnail pane on the bottom of the screen, and an overhead map
is shown on the upper-right side of the screen. Right: a view looking down on the Prague dataset,
rendered in a non-photorealistic style.
5.1 Rendering scenes and transitions in Photo Tourism
This chapter describes the visual user interface and rendering components of the Photo Tourism
system. Figure 5.1 (left-hand side) shows a screenshot from the user interface. The elements of
this window are the main view into the scene, which lls the window, and three overlay panes: an
information and search pane on the left, a thumbnail pane along the bottom, and a map pane in the
upper-right corner.
The main view shows the world as seen from a virtual camera controlled by the user. This view
is not meant to show a photo-realistic rendering of the scene (unless the user is situated directly at a
photo), but rather to display photographs in spatial context and give a sense of the geometry of the
true scene.
The information pane appears when the user visits a photograph. This pane displays information
about that photo, including its name, the name of the photographer, and the date and time when it
was taken. In addition, this pane provides access to geometric controls for searching for other
photographs with certain relations to the current photo.
The thumbnail pane shows the results of search operations as a lmstrip of thumbnails. When
102
the user mouses over a thumbnail, the corresponding image I
j
is projected onto Plane(C
j
) to show
the content of that image and how it is situated in space. The thumbnail panel also has controls for
sorting the current thumbnails by date and time and viewing them as a slideshow.
Finally, when a reconstruction has been aligned to a map, the map pane displays an overhead
view of the scene. The map tracks users position and heading as they move through the scene.
5.1.1 Rendering the scene
While the main emphasis in Photo Tourism is directing users towards photos, the system also can
render the scene using the point, line, and plane geometry recovered from SfM, as shown in the
right-hand side of Figure 5.1. This scene is rendered by drawing the reconstructed points and lines.
In addition, the interface supports a non-photorealistic rendering mode that provides more attractive
visualizations. This mode creates a washed-out rendering designed to give an impression of scene
appearance and geometry, but is abstract enough to be forgiving of the lack of detailed geometry. To
generate the rendering, the system projects a blurred, semi-transparent version of each image I
j
onto
Plane(C
j
), using alpha blending to combine the projected images together to create a watercolor
effect.
The cameras themselves are rendered as small wire-frame pyramids with a solid back face.
If the user is visiting a camera, the back face of that camera pyramid is texture-mapped with an
opaque, full-resolution version of the photograph, so that the user can see it in detail, as shown
in the left-hand image of Figure 5.1. The back faces of other cameras are texture-mapped with a
low-resolution, semi-transparent thumbnail of the corresponding photo.
5.1.2 Transitions between photographs
Another important component of Photo Tourism is the ability to render transitions between photos.
While traditional 2D photo browsers typically use simple cuts or cross-fades to transition between
two photos, the accurate geometric information provided by SfM allows Photo Tourism to use 3D
camera motion and view interpolation techniques to make transitions more visually compelling and
to emphasize the spatial relationships between photographs.
103
Camera motion. During a transitions between two cameras C
start
and C
end
, the systemmoves the
virtual camera along a path between the endpoints (linearly intepolating the locations and rotations,
represented using quaternions). The eld of view of the virtual camera is also interpolated so that
when the camera reaches its destination, the destination image will ll as much of the screen as
possible, while still remaining wholly within the eld of view. The timing of the camera motion is
non-uniform, easing in and out of the transition to maintain second-order continuity [83].
If the camera moves as the result of an object selection (as described in Section 5.2.3), the
transition is slightly different, and tries to keep the object of interest centered in the eld of view.
Before it begins moving, the virtual camera is rst oriented to point at the mean of the selected
points. As the position of the camera is then interpolated towards the destination image, the camera
orientation is updated so as to keep the selected object xed in the center of the view. The nal
orientation and focal length are computed so that the selected object is centered and lls the screen.
View interpolation. While the camera is moving during a transition, the system also interpolates
or morphs between the two photos to create in-between views; this interpolation smooths the visual
transition between the two images. To do the interpolation, I rst compute projection surfaces onto
which the images will be projected during the transition. Then, as the camera moves from C
start
to C
end
, image I
start
is slowly faded out as I
end
is faded in, both projected onto their respective
projection surface. One way to imagine the process is to think of the cameras C
start
and C
end
as
two slide projectors, each projecting onto a screen placed in the scene. During the camera motion,
the rst projector is slowly turned off, as the second is turned on. If the projection surfaces are
close to the actual scene geometry, the effect of this transition is to visually align the two images as
the cross-fade occurs. Combined with the camera motion, this alignment can create a strong sense
of both the physical and visual relationships between the two views. The effect can also be quite
visually striking, especially when morphing between photos with very different appearances, such
as daytime and nighttime.
Photo Tourism supports two types of projection surfaces: (1) planes and (2) polygonal meshes
created from triangulating the point cloud. When planar projection surfaces are used, a single com-
mon plane is computed for the two images used in the transition. When polygonal meshes, two
meshes are used for each transition, one for each image. Figure 5.2 illustrates how transitions work
104
when planar projection surfaces are used.
Planar projection surfaces. In this mode, CommonPlane(C
start
, C
end
) is used as a com-
mon projection surface for the two images. Though a plane is a very simple geometric object and
this method tends to stabilize only a dominant plane in the scene, this technique can still produce
compelling transitions, particularly for urban scenes, which are often composed of planar surfaces.
Polygonal meshes. In this mode, a triangulated mesh is created for each of the two images in-
volved in the transition. To create the meshes, I rst compute a 2D Delaunay triangulation for image
I
start
from the set of 2D projections of Points(C
start
) into I
start
. The projections of Lines(C
start
)
into I
start
are also imposed as edge constraints on the triangulation [22]. The resulting constrained
Delaunay triangulation may not cover the entire image, so I overlay a grid onto the image and add
to the triangulation each grid point not contained inside the original triangulation. Each added grid
point is associated with a 3D point on Plane(C
start
). Thus, we now have a set of triangulated 2D
points that covers the image plane, and a corresponding 3D point for each 2D point. An example
triangulation is shown in Figure 5.4. The 2D triangulation is then used to create a 3D mesh, by
triangulating the 3D points using the same connectivity as their 2D counterparts. This mesh is used
as the projection surface for I
start
. I compute a second projection surface for I
end
in an analogous
way.
During the transition, each of the two images is projected onto its own projection surface. The
depth buffer is turned off to avoid popping due to intersections between the two projection surfaces.
While the meshes are fairly coarse and are not a completely accurate representation of the scene
geometry, they are often suitable as projection surfaces and sufcient to give a sense of the 3D
geometry of the scene. For instance, this approach works well for many transitions in the Great
Wall data set. However, missing geometry, spurious points, and the projection of people and other
foreground elements onto the mesh can often cause distracting artifacts. In my experience, the
artifacts resulting from planar projection surfaces are usually less objectionable than those resulting
from meshes, perhaps because we are accustomed to the distortions caused by viewing planes from
different angles from our experiences in everyday life. Thus, planar impostors are used as the default
projection surfaces for transitions. Example morphs using both techniques are shown in the video
105
(a) (b)
(c) (d) (e)
Figure 5.2: Transitions between images. This gure illustrates how Photo Tourism renders transi-
tions between images (using a planar projection surface). (a) Given two registered images A and B
and the recovered point cloud, a common plane is computed for the image pair by (b) nding the
points visible to both images (colored red), and tting a plane, denoted CommonPlane(A, B) to
these points. For architectural scenes, such planes tends to coincide with walls of buildings, which
are often nearly planar. Parts (c)-(e) show three moments in time during the transition from A to
B. The system starts with the virtual camera at the location of image A (c), and treats A and B as
slide projectors projecting onto CommonPlane(A, B). Initially, projector A is turned completely
on, and projector B is completely off. The system then moves the virtual camera fromAto B, while
turning off projector A and turning off projector B. (d) Midway through the transition, the virtual
camera is halfway between A and B, and both projectors are projecting at 50% opacity, hence the
projected image is a blend of A and B. (e) The end of the transition, with the virtual camera at
B and projector B turned completely on. If the plane is close to the plane of the front wall of the
house, that part of the scene will be stabilized during the transition.
106
(a) (b) (c)
Figure 5.3: Frames from an example image transition. Images (a) and (c) are two photos from the
Prague data set. Image (b) is the midpoint of a transition between (a) and (c), rendered using a
planar projection surface. In this case, the building on the right in image (a) is stabilized.
Figure 5.4: Delaunay triangulation of an image. Left: an image from the Great Wall data set.
Right: Delaunay triangulation of the points seen by the image, augmented with points placed on a
coarse grid over the image.
107
on the project website [108].
Special cases for planar projection surfaces. When planar projection surfaces are used, there
are a few situations that must be handled as special cases. First, if C
start
and C
end
observe no
common points, CommonPlane(C
start
, C
end
) is undened, and the system has no basis for creating
a projection surface for the images. In this situation, the system reverts to rendering the transition
using points, lines, and textured planes, as in Figure 5.1(b), rather than using the images.
1
Second, if the normal to CommonPlane(C
start
, C
end
) is nearly perpendicular to the viewing
directions of C
start
or C
end
, the projected images undergo signicant distortion during the morph.
In this case, I instead use a plane which passes through the mean of Points(C
start
) Points(C
end
),
whose normal is the average of the viewing directions of the two cameras. Finally, if the vanishing
line of CommonPlane(C
start
, C
end
) is visible in either I
start
or I
end
, it is impossible to project the
entirety of I
start
or I
end
onto the plane; imagine turning on a slide projector in an innite room, and
observing how much of the slide projects onto the oorthere is a problem because the horizon of
the room (the vanishing line of the oor) is visible to the slide projector. In this case, I project as
much as possible of I
start
and I
end
onto the plane, and project the rest onto the plane at innity.
5.2 Navigation in Photo Tourism
Photo Tourism supports several different types of controls for navigating through the scene and
nding interesting photographs. These include free-ight navigation, geometric and object-based
search tools, and a stabilized slideshow viewer.
5.2.1 Free-ight navigation
The free-ight navigation controls include standard controls for 3D motion found in games and 3D
viewers. The user can move the virtual camera forward, back, left, right, up, and down, and can
control pan, tilt, and zoom. This allows the user to freely move around the scene and provides a
simple way to nd interesting viewpoints and nearby photographs.
At any time, the user can click on a camera pyramid in the main view, and the virtual camera
1
Chapter 6 introduces a path planning algorithm that gets around this problem by using multi-image transitions.
108
will smoothly move until it is coincident with the selected camera. The virtual camera pans and
zooms so that the selected image lls as much of the main view as possible.
5.2.2 Selecting related views
When visiting a photograph, the user has a snapshot of the world from a single point of view and
a single instant in time. The user can pan and zoom to explore the photo, but may also want to see
aspects of the scene beyond those captured in a single picture. He or she might wonder, for instance,
what lies just outside the eld of view, or to the left of the objects in the photo, or what the view
looks like at a different time of day.
To make it easier to nd such related views, the interface provides the user with a set of geometric
browsing controls. These controls nd photos with certain spatial relationships, and fall into two
categories: zooming controls for selecting the scale at which to view the scene, and directional
controls for seeing more in a particular direction (e.g., left or right). Icons associated with these
controls appear in two rows in the information pane that appears when the user is visiting a photo.
Zooming controls. Photo Tourism supports three zooming controls: (1) nding details, or higher-
resolution close-ups, of the current photo, (2) nding similar photos, and (3) nding zoom-outs, or
photos that show more surrounding context.
Let C
curr
be the photo the user is currently visiting. How can we identify photos that depict
the current view at different scales? One approach would be to nd photos taken from a similar
location, but with different zoom settings. However, the intent of these controls are to show the
content of the image at different levels of detail, rather than to keep the view itself absolutely xed.
For instance, an image taken twenty meters backwards from C
curr
with the same zoom setting may
be just as good a zoom-out as an image taken at the same location as C
curr
with a wider eld of
view. Similarly, a detail of part of the scene may have been taken at a slightly different angle from
C
curr
, in order to capture a more representative or visually pleasing close-up of an object. Therefore,
the zooming relations are based on the content of the photos rather than their absolute positions and
orientations.
To compute these relations, I rst dene the notion of apparent size. Roughly speaking, the
apparent size of an object in an image is the proportion of the image occupied by that object. For
109
instance, a given object might take up much more of an image I
j
than an image I
k
, and thus have a
larger apparent size in I
j
. Such an image could reasonably be called a detail of I
k
.
As the scene is not really divided up into specic objects, I simply use the visible point set
Points(C) as the object visible to camera C and estimate the apparent size of the point set in an
image by projecting it into that image and computing the area occupied by the projections. Specif-
ically, to estimate the apparent size of a point set P in a camera C, I project the points into C,
compute the bounding box of the projections that land inside the image, and calculate the ratio
of the area of the bounding box (in pixels) to the area of the image. I refer to this quantity as
ApparentSize(P, C).
Clicking on one of the scaling controls activates a search for images with the selected relation
(detail, similar, or zoom-out). The images returned from this search are selected from among images
with sufcient overlap with C
curr
(I use images which have at least three points in common with
C
curr
). I classify each such camera C
j
as:
a detail of C
curr
if
ApparentSize(Points(C
j
), C
curr
) < 0.75 ApparentSize(Points(C
j
), C
j
), (5.1)
i.e., Points(C
j
) appears at least 25% larger in C
j
than in C
curr
. In addition, 95% of the points
in Points(C
j
) must project inside the eld of view of C
curr
.
similar to C
curr
if
0.75 <
ApparentSize(Points(C
curr
), C
j
)
ApparentSize(Points(C
curr
), C
curr
)
< 1.3 (5.2)
and the angle between the viewing directions of C
curr
and C
j
is less than a threshold of 10

a zoom-out of C
curr
if C
curr
is a detail of C
j
.
The results of these searches are displayed in the thumbnail strip (sorted by decreasing apparent
size, in the case of details and zoom-outs, i.e., the most zoomed-in view is rst). These controls
are useful for viewing the scene in more detail, comparing similar views of an object which differ
110
Figure 5.5: Zooming controls. Left: an image from the Notre Dame data set. Right: the results
of searching for details (top), similar images (middle), and zoom-outs (bottom), starting from this
input image.
in other respects, such as time of day, season, and year, and for stepping back to see more of the
scene. Results returned from these zooming controls are shown in Figure 5.2.2.
Directional controls. The directional tools give the user a simple way to step left or right, i.e., to
see more of the scene in a particular direction. For each camera, I compute a left and right neighbor
and link these neighbors to arrows displayed in the information pane. To nd the left and right
neighbors of a camera C
j
, I look for images in which the objects visible in C
j
appear to have moved
right or left, respectively.
2
In particular, I compute the average 2D motion m
jk
of the projections of
Points(C
j
) from image I
j
to each neighboring image I
k
:
m
jk
=
1
[Points(C
j
)[

XPoints(C
j
)
(P(C
k
, X) P(C
j
, X)) (5.3)
where P(C, x) is the projection equation dened in Chapter 3. If the angle between m
jk
and the
desired direction (i.e., left = (1, 0)
T
or right = (1, 0)
T
) is less than 15

, and the apparent size of

Points(C
j
) in both images is similar, C
k
is a candidate left (or right) neighbor to C
j
. Out of all
the candidates (if there are any), the algorithm selects the left or right neighbor to be the camera C
k
whose motion magnitude [[m
jk
[[ is closest to 20% of the width of image I
j
.
2
Facing an object and stepping to the right induces a leftward motion of the object in the eld of view.
111
Figure 5.6: Object-based navigation. Left: the user drags a rectangle around the statue in the current
photo. Right: the system nds a new, high-resolution photograph of the selected object, and moves
to that view. Additional images of the object are shown in the thumbnail pane at the bottom of the
screen.
5.2.3 Object-based navigation
Another type of search query supported by the viewer is show me photos of this object, where
the object in question can be directly selected in a photograph or the point cloud. This type of
search, similar to that applied to video in the Video Google system [133], is complementary to, and
has certain advantages over, other types of queries such as keyword search for nding images of a
specic object. The ability to select an object is especially useful when exploring a scenewhen
the user comes across an interesting object, direct selection is an intuitive way to nd a better picture
of that object, and does not require knowing the name of that object.
The user selects an object by dragging a 2D box around a region of the current photo or the
point cloud. All points whose projections lie inside the box form the set of selected points, S. The
system then searches for the best photo of S by scoring each image in the database based on how
well it is estimated to depict S. The top scoring photo is chosen as the representative view, and the
virtual camera is moved to that image. Other images with scores above a threshold are displayed in
the thumbnail strip, sorted by descending score. An example object selection interaction is shown
in Figure 5.6.
My function for scoring images is based on three criteria: (1) whether most of the points in
S are visible in an image, (2) the angle from which the points in S are viewed (frontal views are
112
preferred), and (3) the image resolution. For each image I
j
, I compute the score as a weighted sum
of three terms, E
visible
, E
angle
, and E
detail
.
To compute the visibility term E
visible
, I rst check whether S Points(C
j
) is empty. If so, the
object is deemed not to be visible to C
j
at all, and E
visible
= . Otherwise, E
visible
=
n
inside
|S|
,
where n
inside
denotes the number of points in S that project inside the boundary of image I
j
.
The termE
angle
is used to favor head-on views of a set of points over oblique views. To compute
E
angle
, I rst t a plane to the points in S using a RANSAC procedure. If the percentage of points
in S which are inliers to the recovered plane is above a threshold of 50% (i.e., there appears to be a
dominant plane in the selection), I favor cameras that view the object head-on by setting E
angle
=
V (C
j
) n, where V indicates viewing direction, and n the normal to the recovered plane. If fewer
than 50% of the points t the plane, then E
angle
= 0.
Finally, E
detail
favors high-resolution views of the object. E
detail
is dened to be the area,
in pixels, of the bounding box of the projections of S into image I
j
(considering only points that
project inside the boundary of I
j
). E
detail
is normalized by the area of the largest such bounding
box, so the highest resolution available view will have a score of 1.0.
Once the three scores E
visible
, E
angle
, and E
detail
have been computed for an image, the nal
score is computed as a weighted combination of the three: E = E
visible
+ E
angle
+ E
detail
. In
my system, I set to 1 and to 3. The image that achieves the highest score is selected as the next
image, and a transition to the image is played.
The set S can sometimes contain points that the user did not intend to select, especially oc-
cluded points that happen to project inside the selection rectangle. These extra points can belong to
completely unintended objects, and can therefore cause unexpected images to be returned from an
object selection search. If the system had complete knowledge of visibility, it could cull such hidden
points. However, the fact that the system uses only sparse geometry means that occlusions cannot
be computed without additional assumptions, as innitesimal points will occlude each other with
zero probability. Therefore, I use a set of heuristics to prune the selection. If the selection was made
while visiting an image I
j
, I can use the points that are known to be visible from that viewpoint
(i.e., Points(C
j
)) to rene the selection. In particular, I compute the 3 3 covariance matrix for the
points in S

= S Points(C
j
), and remove from S all points with a Mahalanobis distance greater
than 1.2 from the centroid of S

. If the selection was made while not visiting an image, I instead

113
Figure 5.7: A stabilized slideshow. The user rst nds a photo they like (left-most image), then nds
similar views using the toolbar. The next two images are part of a stabilized slideshow, where Half
Dome is in the same place in each view.
compute a weighted centroid and covariance matrix for the entire set S. The weighting favors points
that are closer to the virtual camera, since these are more likely to be unoccluded than points that are
far away. Thus, the weight for each point is computed as the inverse of its distance from the virtual
camera.
5.2.4 Creating stabilized slideshows
Whenever a search is activated, the resulting images appear in the thumbnail pane; when the thumb-
nail pane contains more than one image, its contents can be viewed as a slideshow by pressing a play
button in the pane. The default slideshow behavior is that the virtual camera will move in 3D from
camera to camera, pausing at each image for a few seconds before proceeding to the next. However,
the user can also choose to lock the virtual camera, xing it to its current position, orientation, and
eld of view. When the images in the thumbnail pane are all taken from approximately the same
location, this locked camera mode stabilizes the images, making it easier to compare one image to
the next. This mode is useful for studying changes in the appearance of a scene over time of day,
seasons, years, weather patterns, etc. An example stabilized slideshow from the Yosemite data set
is shown in Figure 5.7, and in the companion video [108].
5.3 Augmentation and enhancement of scenes
Photo Tourism also allows users to add content to a scene in two ways: a user can (1) register new
photographs to an existing scene and (2) annotate regions of an image and have these annotations
114
automatically propagate to other images.
5.3.1 Registering new photographs
New photographs can be registered on the y to an existing reconstruction using two different tech-
niques, one interactive, the other automatic. In the interactive technique, the user switches to a mode
where an overhead map lls the view, opens a set of images, which are displayed in the thumbnail
panel, and drags and drops each image onto its approximate location on the map. After each image
has been placed on the map, the system estimates the location, orientation, and focal length of each
new photo by running an abbreviated version of the SfM pipeline described in Chapter 3. First, SIFT
keypoints are extracted from the image and matched to the keypoints of the twenty registered images
closest to the specied initial location. The matches to each of the other images are then pruned to
contain geometrically consistent matches, and the 3D points corresponding to the matched features
in the existing registered images are identied. Finally, these matches are used to rene the pose
of the new camera using bundle adjustment. Once a set of photos has been placed on the map, it
generally takes around ten seconds to compute the nal registration for each new camera on my test
machine, a 3.80GHz Intel Pentium 4.
The second way to register a new image requires no initialization. The SIFT features in the new
image are matched directly to the point cloud, and these matches are used for registration. As a
pre-process, a representative SIFT feature descriptor is computed for each 3D point in the scene by
averaging the feature descriptors associated with that point. When a new image is to be registered,
its SIFT keypoints are extracted and matched to the representative feature descriptors. The resulting
matches to 3D points are used to initialize the pose of the new camera by running a direct linear
transform (DLT) procedure [68], and the pose is rened with a local bundle adjustment. Although
this method has the advantage of not requiring initialization, it is slower for very complex scenes and
tends to identify fewer matches than the rst method, and can therefore be somewhat less reliable.
An example registration of a new image to an existing scene (using this second method) is shown in
Figure 5.8.
115
Figure 5.8: A registered historical photo. Left: Moon and Half Dome, 1960. Photograph by Ansel
Adams. This historical photo taken by Ansel Adams was automatically registered to our Half Dome
reconstruction. Right: rendering of DEM data for Half Dome from where Ansel Adams was stand-
ing, as estimated by the system. The white border was drawn manually for clarity. (DEM and color
texture courtesy of the U.S. Geological Survey.)
5.3.2 Creating and transferring annotations
Annotations on image regions are an increasingly popular feature in photo organizing tools such as
Flickr. A unique capability of Photo Tourism is that annotations can be automatically transferred
from one image to all other images that contain the same scene region.
In the Photo Tourism interface, the user can select a region of an image and enter a label for that
region. This annotation is then stored, along with the 3D points S
ann
that lie in the selected area.
When the user visits the photo, the annotation appears as a semi-transparent box around the selected
points. Once annotated, an object can also be linked to other sources of information, such as web
sites, guidebooks, and video and audio clips.
When an annotation is created, it is automatically transferred to all other photographs that see
the annotated object. To transfer an annotation to another image I
j
, the system rst checks whether
the annotation is visible in I
j
, and whether it is at an appropriate scale for the imagethat it neither
lls too much of the image nor labels a very small region. To determine visibility, I simply test that
at least one of the annotated points S
ann
is in Points(C
j
). To check whether the annotation is at an
appropriate scale, I compute the apparent size, ApparentSize(S
ann
, C
j
), of the annotation in image
116
Figure 5.9: Example of annotation transfer. Three regions were annotated in the photograph on the
left; the annotations were automatically transferred to the other photographs, a few of which are
shown on the right. Photo Tourism can handle partial and full occlusions.
I
j
; if the annotation is visible and
0.05 < ApparentSize(S
ann
, C
j
) < 0.8, (5.4)
i.e., the annotation covers between 5 and 80% of the image, the system transfers the annotation to
C
j
. When the user visits C
j
, the annotation is displayed as a box around the annotated points, as
shown in Figure 5.9.
Besides quickly enhancing a scene with semantic information, the ability to transfer annotations
has several applications. First, it enables a system in which a tourist can take a photo (e.g., from a
camera phone that runs my software) and instantly see information about objects in the scene super-
imposed on the image. In combination with a head-mounted display, such a capability could offer
a highly portable, computer-vision-based augmented reality system [44]. Second, it makes labeling
photographs in preparation for keyword search much more efcient; many images can be labeled
with a single annotation.
There are several existing sources of annotated imagery that can be leveraged by my system.
Flickr, for instance, allows users to attach notes to rectangular regions of photos. Tools such as
the ESP Game [155] and LabelMe [123] encourage users to label images on the web and have
accumulated large databases of annotations. By registering such labeled images with an existing
117
collection of photos using our system, Photo Tourism could transfer the existing labels to every
other relevant photo in the system. Other images on the web are implicitly annotated: for instance,
an image on a Wikipedia page is annotated with the URL and topic of that page. By registering
such images to a reconstruction, I could automatically link other photos in that collection to the
same page.
5.4 Results
I have evaluated the features of Photo Tourism on several photo collections. Because Photo Tourism
is an interactive system, the results are best demonstrated through the video captures of interactive
sessions found on the Photo Tourism project website [108]. Screenshots from these sessions are
shown in gures throughout this chapter.
The rst two collections were taken under relatively controlled settings (i.e., a single person with
a single camera and lens): Prague, a set of 197 photographs of the Old Town Square in Prague,
Czech Republic, taken over the course two days, and Great Wall, a set of 120 photographs taken
along the Great Wall of China (82 of which were ultimately registered). The images in the Prague
data set were taken by a photographer walking along the sides of the square, facing the buildings. A
few more distant views of the buildings were also captured. The directional and zooming controls
work particularly well on this collection, as demonstrated in the companion video. In addition, the
reconstruction was registered to an aerial image, and therefore a map appears as part of the interface.
For the Great Wall collection, one interesting way to view the images is to play them in order as a
slideshow, simulating the effect of walking along the wall in the photographers footsteps. For this
collection, the mesh-based projection surfaces are particularly effective.
I also demonstrate the system on several uncontrolled sets consisting of images downloaded
from Flickr. Trevi Fountain is a set of 360 photos of the Trevi Fountain in Rome, registered from
466 photos matching trevi AND rome. Notre Dame is a set of 597 photos of the front facade
of the Notre Dame Cathedral in Paris. These photos were registered starting from 2,635 photos
matching the search term notredame AND paris. Yosemite is a set of 325 photos of Half Dome in
Yosemite National Park, registered from 1,882 photos matching halfdome AND yosemite.
The Trevi Fountain collection contains several interesting, detailed sculptures. The video seg-
118
ment for this collection demonstrates the object selection feature, in which the system nds and
transitions to a good photo of an object selected by the user, as shown in Figure 5.6. Notre Dame
contains photos at many different levels of detail, from images taken from over a hundred meters
away from the cathedral, to extreme close-ups showing intricate details of the sculptures above the
central portal. The total difference in scale between the extreme views is a factor of more than
1,000. Thus, the zooming features of Photo Tourism are well-suited to this collection. The video
also demonstrates the annotation transfer feature on this collection, as well as the selection of an-
notations for display based on the scale of the current image. Finally, the Half Dome collection
contains many spectacular images of the mountain during different seasons and times of day. The
stabilized slideshow feature, demonstrated in the video and in Figure 5.7, is an effective way to view
the resultant variation in the appearance of the mountain.
After the initial reconstruction of the Half Dome collection, I aligned it to a digital elevation
map using the approach described in Chapter 3, Section 3.4.1. I then registered a historical photo,
Ansel Adams Moon and Half Dome, to the data set, using the automatic method described in
Section 5.3.1. The result is an estimate of where Adams was standing when he took the photo.
Figure 5.8 shows a synthetic rendering of Half Dome from this estimated position.
5.5 Discussion
This chapter described Photo Tourism, a new interface for exploring photo collections related by
place. Photo Tourism situates the user in a 3D world among the reconstructed photo collection,
and provides new tools for searching for images, including zooming controls, directional controls,
and an object selection tool. Photo Tourism also enables new ways to visualize and render scenes,
through 3D transitions between images and stabilized slideshows. Finally, it provides a powerful
annotation mechanism that can be used to label an entire collection of images at once.
Photo Tourism also has several limitations. The 3D transitions work well for images of planar
or near-planar parts of a scene, such as walls. The results can be less attractive for images that view
more complex 3Dgeometry, or for images that viewmultiple dominant planes. The mesh-based pro-
jection surfaces can handle non-planar geometry, but produce their own artifacts, especially where
the geometry is inaccurate or unmodeled, or when people or other foreground objects are projected
119
onto the mesh. More accurate geometry produced, e.g., from a laser scan or with multi-view stereo,
would solve some, but not all, of these problems. Hence, better methods for rendering transitions
between photos containing complex surfaces and foreground objects is an important open research
challenge.
The navigation controls provided by Photo Tourism are primarily photo-centric, designed to
move the user from one photo to another. While these can be effective tools for exploring the photo
collection, they are not necessarily the best interface for exploring the scene itself. A more common
approach, used in most 3D games, is to give users continuous control over their movement, rather
than constraining them to a particular xed set of views. Of course, in Photo Tourism, users can
choose to use the free-viewpoint navigation controls, but they then lose the realistic views afforded
by the images. In any case, it is not clear that even free-viewpoint navigation controls are the ideal
interface for exploring scenes, as they can provide too many degrees of freedom, leaving a user
with too many possible directions to explore. In the next chapter, I present a second user interface
designed for exploring scenes. The interface analyzes the distribution of camera viewpoints in large
photo collections to create sets of constrained 3D navigation controls tailored to specic locations.
120
Chapter 6
FINDING PATHS THROUGH THE WORLDS PHOTOS
The previous chapter described Photo Tourism, a 3D photo browser that provides new geomet-
ric controls for exploring photo collections. Let us now consider a slightly different task: explor-
ing scenes. How well does Photo Tourism work as a tool for exploring a large space such as the
Prague Old Town Square? As demonstrated in the previous chapter and accompanying video, Photo
Tourism allows us to move around the square in several ways. We can zoom in or out of a photo
to see more detail or more context. We can move to the left or right to see a photo of the next
building over in a row of shops. We can also drag a box around an interesting object or building,
and transition to a good photo of that selected object.
All of these controls are potentially useful, but what are the best possible controls? To attempt
to answer that question, lets step back for a moment, and consider what we want to accomplish
in exploring the scene. Suppose we have never been to the square. We might know that it is a
signicant place, with interesting sights, but we may not know exactly where we should go or what
we should see. Indeed, the square contains several notable landmarks, including the Old Town
Hall with its astronomical clock, the T` yn Cathedral, the St. Nicholas Church, and the statue of Jan
Hus. An ideal 3D navigation interface might make it easy to nd all of these landmarks, and offer
additional controls, such as a way to orbit the Jan Hus statue and see it from different angles, when
suitable. Another scenethe Vatican, saywill have a completely different set of interesting views
(the Sistine Chapel, Michelangelos Piet` a, St. Peters Square) and might require a different set of
controls. Scenes also come in a wide variety of shapes and sizes. The Old Town Square consists of
rows of buildings, while other scenes, such as the Statue of Liberty, consist of a single, dominant
object.
This variability is at the heart of the problem I consider in this chapter: how can we discover
or devise the best navigation controls for any given scene? The solution I present in this chapter is
to derive controls from large Internet photo collections of the scene. These collections of photos,
121
captured by many people, provide an extremely valuable source of information for creating controls
for exploration tasks, as they represent samples of how people actually experienced the scene, where
they stood, and what views they found interesting.
The navigation controls provided by Photo Tourism also adapt to photo collections, but only at
a very local level. For each photo, considered in isolation, a set of possible details, similar images,
left and right neighbors, etc., are computed. Analyzing the structure of the photo collection at this
level might miss important features of the scene. Of course, Photo Tourism also provides standard
continuous, free-viewpoint navigation controls, as do the vast majority of games and simulators,
that let users move wherever they want. However, while such unconstrained controls might make
it possible to explore the scene in any particular way, they still might not make it easy; they may
provide too many degrees of freedom when only a few are required (e.g., the one-dimensional
case of rotating around an object), and do not directly solve the problem of getting users where they
really want to, or ought to, go. Thus, devising a good interfaces for exploring scenes is a challenging
problem. As image-based rendering (IBR) methods scale up to handle larger and larger scenes, the
problem of devising good viewpoint controls becomes even more critical.
The overarching issue here is that to quickly and efciently explore understand a scene, a user
may want a variety of controlsPhoto Tourism-style controls (for instance, for nding a photo of a
selected object), continuous, object-based navigation controls (to orbit around a statue), automatic
controls (to take the user to an interesting landmark), and perhaps even controls that suggest where
to go next. These various controls depend on the task the user wants to perform, the contents of the
scene (e.g., does it contain objects?), and the part of the scene the user is currently visiting.
One solution to this problem is to manually plan a set of desired paths through a scene and
capture those views, then design a user interface that only exposes controls for following these
paths. This approach is used for many IBR experiences such as panoramas, object movies [21], and
moviemaps [87]. While effective, this kind of approach cannot leverage the vast majority of existing
photos, including the millions of images of important landmarks available through the Internet.
While deriving such controls is a challenging research problem, using Internet photo collections
to generate controls also has major advantages. Internet photo collections represent samples of
views from places people actually stood and thought were worth photographing. Therefore, through
consensus, they tend to capture the interesting views and paths through a scene. I leverage this
122
Figure 6.1: Paths through the Statue of Liberty photo collection. Left: several Flickr images of the
Statue of Liberty, from a collection of 388 input photos. Right: reconstructed camera viewpoints
for this collection, revealing two clear orbits, shown here superimposed on a satellite view. The goal
is to automatically discover such orbits and other paths through view space to create scene-specic
controls for browsing photo collections.
observation to generate controls that lead users to interesting views and along interesting paths.
For example, consider the overhead view of a reconstruction of the Statue of Liberty from 388
photos downloaded from Flickr, shown in Figure 6.1. Most of the photos in this collection were
captured from the island or from boats out in the water, and are distributed roughly along two
circular arcs. Even if we knew absolutely nothing about the content of the scene, this distribution of
views would tell us that (a) there is some very important thing in the scene, since almost all images
are trained on the same point, and (b) the object of interest is typically viewed from a range of
angles forming arcs around the object. The distribution thus suggests two natural orbit controls for
browsing this scene. While the viewpoints in this scene have a particularly simple structure, I have
observed that many Internet collections can be modeled by a combination of simple paths through
the space of captured views.
In this chapter, I describe the Pathnder system,
1
which extends Photo Tourism to provide a
uid, continuous IBR experience with effective scene-specic controls derived from the distribution
of views in a photo collection. As with Photo Tourism, there are both rendering and navigation
aspects to this problem. On the rendering side, I introduce new techniques for selecting and warping
images for display as the user moves around the scene, and for maintaining a consistent scene
1
The work described in this chapter originally appeared at SIGGRAPH 2008 [136]. Rahul Garg contributed signi-
cantly to the ideas and results described in this chapter.
123
appearance. On the navigation side, I describe my approach for automatically generating good
controls from a photo collection and integrating these controls into the user interface. These two
problems, rendering and navigation, are not separate, but are intimately linked together, for two
reasons. First, my approach to rendering is based on a new view scoring function that predicts the
quality of reprojecting an input photo to a new viewpoint. This function is used to select good views
for rendering, but also to nd trajectories through viewpoint space with high visual quality, suitable
for use as navigation controls. Second, the projection surfaces used for rendering are selected based
on how the user is currently moving through the scene.
The remainder of this chapter describes the various components of the Pathnder system. First,
however, I review related work in 3D navigation interfaces to motivate the design of the system.
6.1 Related work on 3D navigation
A primary goal of the Pathnder system is to create controls that make 3D navigation through a
scene easy and intuitive. While computer graphics researchers have concentrated mainly on render-
ing of scenes, there has also been a wealth of research on interfaces for 3D navigation, primarily in
the eld of human-computer interaction. This section surveys previous work on 3D navigation and
distills a set of key guiding principles for the design of navigation interfaces.
6.1.1 Principles of 3D scene navigation
3D scene navigation refers to manipulating a rst-person viewpoint in a virtual 3D scene. It involves
translating some inputfrom a mouse, keyboard, head-mounted display, or other deviceinto a 6D
set of camera parameters which specify the 3Dposition and 3Dorientation of the camera (sometimes
with an additional degree of freedom for zoom). 3D navigation is a surprisingly difcult problem;
without good exploration and visualization tools it can be easy to get lost in a scene [131]. The
development of interfaces for navigating virtual 3D environments dates back at least to the work
of Sutherland in the 1960s [142]. Since then such interfaces have appeared in numerous settings:
games, simulations, 3D modeling software, mapping applications such as Google Earth, virtual
reality systems, and others.
Before considering the design of navigation interfaces, however, it is useful to consider the goals
124
of 3Dnavigation, i.e., what tasks a user may be interested in performing in a 3Denvironment. Tan, et
al., [145], present a task-based taxonomy of navigation controls which enumerates three categories
of tasks:
Exploration: gaining survey knowledge, e.g., building a mental map of the scene.
Search: nding and moving to a particular view, region, or object.
Object inspection: exploring a range of desired views of a particular object.
While this list is not necessarily comprehensive, these tasks are relevant to the kinds of experiences I
seek to enable in Pathnder. The rst of these, exploration, is particularly relevant; my work focuses
on exploration of popular real world scenes, scenes which many users may be interested in visiting
or learning more about, but with which they may not initially be intimately familiar. For such users,
an ideal interface might be a set of simplied controls that make it easy to nd the interesting parts
of the scene.
What kind of controls best aid a user in accomplishing these tasks? While there is no simple
answer, certain principles for 3D interface design can be found in the literature on 3D navigation.
The right navigation controls are scene- and task-dependent. Researchers have found that nav-
igation controls that are entirely natural in one setting can be completely ineffective in another. In
one study, Ware and Osbourne [157] evaluated several types of navigation controls, including ying
vehicle, where a user moves around the scene by directly moving and turning the virtual camera,
and scene-in-hand, where the user instead moves and rotates the scene itself. They found that in
scenes consisting of a single object, test subjects overwhelmingly found the scene-in-hand controls
to be most natural and the ying vehicle controls to be the least natural, while the exact opposite
was true for complex scenes such as a maze. Similarly, different types of controls have been shown
to be more useful depending on the task a user is trying to perform, such as getting an overview of
a scene or inspecting a particular object [145].
For these reasons, many 3D interfaces, including those used in the Maya 3D modeling software
and several VRML browsers, provide different sets of controls which users can toggle between,
such as camera-centric and object-centric controls. However, manually toggling between modes
may be more cumbersome than necessary. Another approach is to learn the right set of controls for
125
a given scene, and a few researchers have explored this idea. Singh and Balakrishnan [131] propose
a system that attempts to learn good visualizations of a scene based on a set of previous interactions
with the same scene. In the Pathnder system, such controls are learned from the distribution of
views in a large photo collection.
Constrained navigation can be more efcient. A common problem with 3D viewers that allow
completely unrestricted motion is that it can be easy to get lost in space, getting in a state where the
user doesnt know where they are or how to get back. Even if a user stays grounded, unconstrained
controls may provide more degrees of freedom than are strictly necessary, making them more dif-
cult to use. Furthermore, common 2D input devices, such as a mouse, cannot always be naturally
mapped to the 6 or 7D space of camera views. Constraining the navigation controls to a smaller
number of degrees of freedom is one way to solve these problems. Indeed, some approaches, such
as Galyeans River Analogy [50], simply move users automatically along a pre-specied path, but
give the user freedom to control certain parameters, such as viewing direction and speed. Hanson, et
al. [64] suggest several design techniques for adding such constraints to a controller, such as making
the orientation of the camera dependent on the position. They also present a preliminary evaluation
showing that adding constraints can help improve a users survey knowledge about an environment.
Several specic constrained navigation techniques have been developed, such as Hovercam [78],
which converts 2D input to a generalized orbit around an arbitrary object for tasks that require close
object inspection. The main drawback of constrained navigation is that it can give a user less of a
sense of freedom [64].
Use high-level goals and automatic navigation when possible. Often, automatic navigation is
the easiest way for a user to satisfy a certain goal [63, 36, 38]. For instance, suppose a user is
virtually visiting Prague and has the goal of nding the Jan Hus Memorial. It would probably
be much easier for the user to express that goal to the system, and have the system automatically
move the user to the Memorial, than for the user to try and hunt for it himself. A good example
of a system that implements this behavior is Google Earth, which supports automatic navigation
through its search features. In Google Earth, a user can search for a place by typing in a name or
address, and a smooth camera motion automatically brings the user to the destination. The path
126
along which the user is transported is a factor in how comprehensible the transition is. The quickest
path from point A to point B is teleportationinstantaneously cutting from one view to the next;
however, teleportation can be quite disorienting [14]. Another approach, explored by Santos, et al.
[36], moves the user along shortest paths in 3D space between two views. Drucker and Zeltzer [38]
describe a general system where multiple tasks and constraints can be combined in a path planning
algorithm to generate automatic, natural camera motions.
The users spatial position and motion should always be clear. Unless users can easily tell
where they are and how they are moving, they can easily become disoriented [63]. If the scene
has a variety of visual content, then landmarks assist users in keeping track of their position [52].
Other, more explicit cues have also been investigated, such as augmenting a scene with gridlines or
additional landmarks [28]. One particular cuean overhead map which displays or tracks the users
positionis very commonly used [28, 152], and has been shown to substantially aid in navigation
tasks [27].
These interface considerations are all relevant to the design of the Pathnder system. However,
Pathnder takes these ideas a step further by trying to automatically infer good sets of navigation
controls by analyzing how many different people have photographed a given scene. In particular,
Pathnder nds scene-specic, constrained navigation controls, such as orbits around objects, and
suggests these controls to the user. In addition to these controls, Pathnder attempts to nd im-
portant views of the scene, and provides automatic controls for moving to these important views.
Paths between these views are generated with a new path planning algorithm that attempts to create
understandable paths by mimicking how people move through the scene.
6.2 System overview
The Pathnder system takes as input a set of photos from a variety of viewpoints, directions, and
conditions, taken with different cameras, and with potentially many different foreground people and
objects. From this input, the systems creates an interactive 3D browsing experience in which the
scene is depicted through photographs that are registered and displayed as a function of the current
viewpoint. The system guides the user through the scene by means of a set of automatically com-
127
puted controls that expose detected orbits, panoramas, interesting views, and optimal trajectories
specic to the scene and distribution of input views.
The system consists of the following components:
A set of input images and camera viewpoints. The input is an unstructured collection of pho-
tographs taken by one or more photographers. We register the images using the SfM algorithms
described in Chapters 3 and 4.
Image reprojection and viewpoint scoring functions that evaluate the expected quality of render-
ing each input image at any possible camera viewpoint. The reprojection process takes into account
such factors as viewpoint, eld of view, resolution, and image appearance to synthesize high quality
rendered views. The viewpoint scoring function can assess the quality of any possible rendered
view, providing a basis for planning optimal paths and controls through viewpoint space.
Navigation controls for a scene. Given the distribution of viewpoints in the input camera database
and the viewpoint scoring function, the system automatically discovers scene-specic navigation
controls such as orbits, panoramas, and representative images, and plans optimal paths between
images.
A rendering engine for displaying input photos. As the user moves through the scene, the ren-
dering engine computes the best scoring input image and reprojects it into the new viewpoint, trans-
formed geometrically and photometrically to correct for variations between images. To this end, I
introduce an orbit stabilization technique for geometrically registering images to synthesize motion
on a sphere, and an appearance stabilization technique for reducing appearance variation between
views.
A user interface for exploring the scene. A 3D viewer exposes the derived controls to users,
allowing them to explore the scene using these controls, move between different parts of the scene,
or simply y around using traditional free-viewpoint navigation. These controls combine in an
intuitive way.
These components are used in the three main stages of the Pathnder system. First, the ofine
structure from motion process recovers the 3D location of each photograph. Next, the scene is au-
tomatically processed to derive controls and optimal paths between images. Finally, this augmented
scene can be browsed in an interactive viewer.
128
The remainder of the chapter is organized as follows. In the next section, I describe the view
scoring function at the heart of the system. Section 6.4 describes how scene-specic navigation
controls are derived from a photo collection and integrated with other modes of navigation. Sec-
tion 6.5 describes the scene rendering engine. Section 6.6 describes the path planning algorithm
used to compute transitions between views. Section 6.7 describes techniques used to stabilize the
appearance of images as the user moves through the scene. Section 6.8 demonstrates the approach
on a variety of scenes and for a range of visualization tasks including free-form 3D scene brows-
ing, object movie creation from Internet photos or video, and enhanced browsing of personal photo
collections. Finally, Section 6.9 discusses limitations of the system and ideas for future work.
6.3 Viewpoint scoring
My approach is based on (1) the ability to reproject input images to synthesize new viewpoints, and
(2) to evaluate the expected quality of such reprojections. The former capability enables rendering,
and the latter is needed for computing controls that move the viewer along high quality paths in
viewpoint space. In this section I describe my approach for evaluating reprojection quality.
As in Photo Tourism, the Pathnder system is given a database of input images 1 whose camera
parameters (intrinsics and extrinsics) have been computed. The term camera denotes the viewing
parameters of an input image I
j
. Each camera C
j
observes a set of points Points(C
j
). The term
image denotes an input photo I
j
from the database. The term viewpoint denotes a set of viewing
parameters v in the virtual scene, and the term view denotes an output photo that we seek to render
from a given viewpoint. A view is produced by reprojecting an input photo, through a rendering
process, to the desired new viewpoint v.
The rst step is to dene a reprojection score S
proj
(I
j
, v) that rates how well a database image I
j
can be used to render a new view at v. The best reprojection is obtained by maximizing S
proj
(I
j
, v)
over the image database, yielding a viewpoint score S
view
(v):
S
view
(v) = max
I
j
I
S
proj
(I
j
, v) (6.1)
Ideally, S
proj
(I
j
, v) would measure the difference between the synthesized view and a real photo of
the same scene captured at v. Because the system does not have access to the real photo (unless v is
129
Figure 6.2: Angular deviation. The angular deviation factor of the viewpoint score penalizes the av-
erage angular deviation between rays from the current viewpoint and the image under consideration
through a set of points in the scene. Here, the angle penalty
ang
is shown for one of the points.
exactly coincident with an input camera), I instead use the following three criteria to estimate how
well image I
j
represents v:
1. Angular deviation: the relative change in viewpoint between C
j
and v should be small.
2. Field of view: the projected image should cover as much of the eld of view of v as possible.
3. Resolution: image I
j
should be of sufcient resolution to avoid blur when projected into v.
For a given image and viewpoint, each of these criteria is scored on a scale from 0 to 1. To compute
these scores, we require a geometric proxy for each image to use as a projection surface during
reprojection into v; the proxy geometry is discussed in Section 6.5.2.
Angular deviation. The angular deviation score S
ang
(I
j
, v) is proportional to the angle between
rays from viewpoint v through a set of points in the scene and rays from camera C
j
through the same
points. This is akin to the minimum angular deviation measure used in Unstructured Lumigraph
Rendering [17]. Rather than scoring individual rays in the database, however, the system scores
entire images by averaging the angular deviation over a set of 3D points observed by I
j
These points,
denoted SamplePoints(C
j
), are selected from Points(C
j
) for each image in a pre-processing step.
To compute SamplePoints(C
j
), I project Points(C
j
) into I
j
, distribute them into a 10 10 grid of
bins dened on the image plane, then select one point from each non-empty bin. This ensures that
the points are relatively evenly distributed over the image.
130
The average angular deviation is computed as:
S

ang
(I
j
, v) =
1
[ SamplePoints(I
j
)[

XSamplePoints(I
j
)
angle(Xc
j
, Xc(v)). (6.2)
where c
j
is the 3D position of camera C
j
, c(v) is the 3D position of viewpoint v, and angle(a, b)
gives the angle between two rays a and b. The average deviation is clamped to a maximum value
of
max
(for all my examples, I set
max
= 12

), and mapped to the interval [0, 1]:

S
ang
(I
j
, v) = 1
min(S

ang
(I
j
, v),
max
)

max
. (6.3)
A score of 1 indicates that C
j
and v are coincident viewpoints, and a score of 0 means that the
average angle between corresponding rays is greater than
max
.
Field-of-view score. The eld-of-view score S
fov
(I
j
, v) is computed by reprojecting I
j
into v and
computing the area of the view at v that is covered by the reprojected image. I compute a weighted
area, with higher weight in the center of the view, as I nd that it is generally more important to
cover the center of the view than the boundaries. The weighted area is computed by dividing the
view into a grid of cells, (, and accumulating weighted contributions from each cell:
S
fov
(I
j
, v) =

G
i
G
w
i
Area(Project(I
j
, v) G
i
)
Area(G
i
)
, (6.4)
where Project(I
j
, v) is the 2D polygon resulting from reprojecting image I
j
into viewpoint v (if
any point of the projected image is behind viewpoint v, Project returns the empty set); Area is the
area, in pixels, of a 2D polygon. The weights w
i
approximate a 2D Gaussian centered at the image
center.
Resolution score. Finally, the resolution score S
res
(I
j
, v) is computed by projecting I
j
into v
and nding the average number of input pixels of I
j
used per output screen pixel. An image of
sufcient resolution should have at least one input pixel per output pixel. This score is thus computed
as the ratio of the number of pixels in I
j
to the area, in screen pixels, of the reprojected image
131
Project(I
j
, v):
S

res
(I
j
, v) =
Area(I
j
)
Area(Project(I
j
, v))
. (6.5)
If this ratio is greater than one, then, on average, the resolution of I
j
is sufcient to avoid blur when
I
j
is projected into the view (the system uses mip-mapping to avoid aliasing, so a ratio greater than
one is acceptable). I transform S

res
to map the interval [ratio
min
, ratio
max
] to [0, 1]:
S
res
(I
j
, v) = clamp
_
S

res
(I
j
, v) ratio
min
ratio
max
ratio
min
, , 1
_
, (6.6)
where clamp(x, a, b) clamps x to the range [a, b]. I use values of 0.2 and 1.0 for ratio
min
and
ratio
max
, and enforce a non-zero minimum resolution score because I favor viewing a low-
resolution image rather than no image at all.
The three scores described above are multiplied to give the nal reprojection score S
proj
:
S
proj
(I
j
, v) = S
ang
(I
j
, v) S
fov
(I
j
, v) S
res
(I
j
, v). (6.7)
Again, to compute the viewpoint score S
view
, the reprojection score is computed for each input
image; S
view
is the maximum over these scores.
6.4 Scene-specic navigation controls
In Section 6.1 I laid out several guidelines for the design of 3D navigation control. Three primary
guidelines are that the right navigation controls are scene- and task-dependent, constrained navi-
gation can be more efcient, and one should use high-level goals and automatic navigation when
possible. Based on these observations, the Pathnder system attempts to (a) derive constrained,
scene-specic controls for different parts of the input scene, (b) nd interesting views that users
might want to visit, and (c) provide automated, optimized paths from getting between any pair of
views or controls.
To these ends, the large Internet photo collections used as input to the system are extremely
useful, in that the tend to cluster along interesting paths through a scene, and around interesting
objects, and thus provide important cues as to what controls should be provided. Of course, the
regions near the input samples will also be the areas where we can likely render good views of the
132
scene (i.e., views for which S
view
(v) is high). I take advantage of this information through a set of
automatic techniques for deriving controls from a reconstructed scene. The result of this analysis
is a set of scene-specic controls. For instance, the Statue of Liberty scene shown in Figure 6.1
has two scene-specic controls: an orbit control for the cluster of views taken on the island, and a
second orbit for views taken out on the water.
In the rest of this section, I outline the navigation modes of the Pathnder system and describe
how scene-specic controls are computed.
6.4.1 Navigation modes
Pathnder supports three basic navigation modes:
1. Free-viewpoint navigation.
2. Constrained navigation using scene-specic controls.
3. Optimized transitions from one part of the scene to another.
Free-viewpoint navigation. The free-viewpoint navigation mode allows a user to move around
the scene using standard 6-DOF (3D translation, pan, tilt, and zoom) ying vehicle navigation
controls. I also provide an orbit control which allows for scene-in-hand-style rotation about an
object.
While free-viewpoint controls give users the freedom to move wherever they choose, they are
not always the easiest controls for move around complex scenes, as the user has to continually
manipulate many degrees of freedomwhile (at least in IBR) ideally staying near the available photos.
Scene-specic controls. Pathnder supports two types of scene-specic controls: orbits and panora-
mas. Each such control is dened by its type (e.g., orbit), a set of viewpoints, and a set of images
associated with that control. For an orbit control, the set of viewpoints is a circular arc of a given
radius centered at and focused on a 3D point; for a panorama the set of viewpoints is a range of
viewing directions from a single 3D nodal point. When a control is active, the user can navigate
through the corresponding set of viewpoints using the mouse or keyboard. In addition to scene-
specic orbits and panoramas, I also compute a set of representative canonical views for a scene.
133
Transitions between controls. The nal type of control is a transition between scene-specic
controls or canonical views. Pathnder allows users to select a control or image they are interested
in exploring next; once a destination is selected, the virtual viewpoint is then moved on an automated
path to the new control or image. The transition is computed using a new path planning algorithm
that adapts to the database images, as described in Section 6.6. This method of directly selecting and
moving to different parts of the scene is designed to make it easy to nd all the interesting views,
following the guideline that automatic navigation is often the most efcient.
6.4.2 Discovering controls
Once a scene is reconstructed, the Pathnder system automatically analyzes the recovered geom-
etry to discover interesting orbits, panoramas, and canonical views. Orbits typically correspond
to interesting objects which can be viewed from many different angles, panoramas are viewpoints
from which the scene is photographed from many different viewing directions (e.g., the top of a tall
building), and canonical views correspond to important viewpoints represented many times in the
database. This section describes how each of these are discovered in a scene.
Orbit detection. I dene an orbit to be a distribution of views positioned on a circle and con-
verging on (looking at) a single point, denoted the convergence point p
focus
. I only consider circles
parallel to the ground plane as valid orbits, as most orbits that appear in the world are around objects
photographed from different directions by people standing on a oor or other plane. I further con-
strain p
focus
to lie on a vertical axis o passing through the center of the circle and perpendicular to
the ground plane. The height of p
focus
determines the tilt at which the object of interest is viewed.
Because full 360

view distributions are uncommon, I allow an orbit to occupy a circular arc. I

dene a good orbit to be on that satises the following objectives, as illustrated in Figure 6.3:
quality: maximize the quality of rendered views everywhere along the arc.
length: prefer arcs that span large angles.
convergence: prefer views oriented towards the center of the orbit.
object-centered: prefer orbits around solid objects (as opposed to empty space).
134
Figure 6.3: Scoring a candidate orbit. An orbit is evaluated by regularly sampling viewpoints along
the candidate arc (here, the arc is shown in gray, and the samples are shown as small circles drawn
on top of the arc). For each sample position, we want to nd a nearby image with a high reprojection
score that is oriented towards the orbit center (thus eliminating the red cameras). The light samples
score well on these objectives while the black samples do not. We search for large arcs where the
sum of the sample scores is high, and that do not contain large low-scoring gaps. Here, the optimal
subset of the arc is shown with the curved black arrow.
Given these objectives, the problem of detecting orbits involves (1) dening a suitable objective
function, (2) enumerating and scoring candidate orbits, and (3) choosing zero or more best-scoring
candidates. One could imagine many possible techniques for each of these steps; in what follows, I
describe one approach that has worked quite well in practice.
I rst dene my objective function for evaluating orbits. An orbit is fully specied by a center
orbit axis o and an image I
j
; the image denes the radius of the circle (the distance of the camera
center from o), and the convergence point p
focus
on the orbit axis (p
focus
is the closest point on the
axis o to the optical axis of camera C
j
). Assume further that C
j
is the point on the arc midway
between the endpoints of the arc. The geometry of such an orbit is illustrated in Figure 6.4.
I dene the orbit scoring function, S
orbit
(o, I
j
), to be the sumof individual viewscores, S
orbit
(o, I, ),
sampled at positions along the arc. To compute S
orbit
(o, I
j
, ) at a sample viewpoint v() (the
viewpoint on the arc at angle fromI
j
), I look for support for that view in the set of database images
1. In particular, I score each image I
k
1 based on (a) how well I
k
can be used to synthesize a
view at v() (estimated using the reprojection score S
proj
(I
k
, v())), and (b) whether I
k
is looking
135
Figure 6.4: An orbit can be dened by an orbit axis and a camera. A vertical orbit axis orbit o,
combined with a camera C
j
(which denes the orbit radius), denes a family of orbit arcs around
the axis. This family consists of all arcs of a circle centered on the axis which passes through the
camera center (and for which the camera center is at the middle of the arc). The focal point p
focus
is the point on the axis o closest to the ray through the center of C
j
.
at the orbit axis (the convergence score). S
orbit
(o, I
j
, ) is then the score of the best image I
k
at
v():
S
orbit
(o, I
j
, ) = max
I
k
I
S(I
k
, v()) f
o
(I
k
). (6.8)
The convergence score f
o
(I
k
) is dened as:
f
o
(I
k
) = max
_
0, 1

max
_
(6.9)
where = angle(v(C
k
), p
focus
p(C
k
)), i.e., the angle between the viewing direction v(C
k
) of
image I
k
and the ray from the optical center p(C
k
) of I
k
to p
focus
; I use a value of
max
= 20

.
This term downweights images for which p
focus
is not near the center of the eld of view.
I place a few additional constraints on the images I
k
considered when computing S
orbit
(o, I
j
, ):
p
focus
must be inside the eld of view and in front of I
k
.
The tilt of I
k
above the ground plane must be less than 45

(orbits with large tilt angles do not

produce attractive results).
There must be a sufcient number (I use k = 100) of 3D points visible to I
k
whose distance
136
from I
k
is less than the orbit radius. I enforce this condition to ensure that we nd orbits
around an object (as opposed to empty space).
I compute S
orbit
(o, I
j
, ) at every degree along the circle 180 < 180. For simplicity, I
refer to these sample scores as s(). A good orbit will have a long arc of relatively high values of
s(). Simply summing the values s(), however, could favor orbits with a few sparsely scattered
good scores. Instead, I explicitly nd a long chain of uninterrupted good scores centered around I
j
,
then add up the scores on this chain. I dene this chain as the longest consecutive interval [
L
,
L
]
such that the maximum s() in each subinterval of width 15

is at least = 0.01. This denition

allows for small gaps, or intervals with low scores, in the arc. If this longest chain subtends an
angle less than 60

, the score of the orbit is zero. Otherwise, the score is the sum of the individual
scores in the chain:
S
orbit
(o, I
j
) =
L

k=L
s(
k
). (6.10)
Nowthat we have an objective function, we need a way to enumerate candidate orbits (mainly for
sake of efciency; we could score all possible orbits). My strategy for nding such orbits operates
in two stages. First, I compute a set of good candidate orbit axes, by nding axes in 3D space that
are the convergence points of many database images. Second, I create and score candidate orbits by
pairing each candidate axis with each database image I
j
.
To identify a set of candidate orbit axes, I take an approach similar to that of Epshtein et al.[40]
and use the idea that an interesting object will often occupy the center of the eld of view of many
images. I rst project all cameras onto the ground plane
2
and compute the 2D intersections of the
optical axes of all pairs of cameras (discarding intersection points which lie in back of either camera,
or which are not approximately equidistant from both cameras). I then compute the density D of
these intersection points at each point x:
D(x) = number of intersection points within distance w of x.
(I use w = 0.02, although this value should ideally depend on the scale of the scene). The local
maxima in this density function identify vertical axes in the scene which are at the center of many
2
This projection, reducing the problem to 2D, is possible because all orbit axes are assumed to be vertical.
137
Figure 6.5: Density function and detected orbits for the Pantheon data set. Top row, left: the set of
intersection points of viewing axes is superimposed on the point cloud of the Pantheon. The color
of each point corresponds to the density of points in its neighborhood (a red point has the highest
density, a dark blue point the lowest). There are several clear maxima of this function, including a
point just behind the front facade and a point at the altar (at the extreme left of the gure).
4
Top row,
right: the three nal detected orbits shown as blue arcs centered at the red orbit points. Bottom row:
images from each of the three detected orbits.
different photos, and are thus potentially interesting. A plot of the density function for the Pan-
theon dataset is shown in Figure 6.5, along with the three detected orbits. These orbits are also
demonstrated in the video accompanying this chapter [107].
Next, I nd the point with the highest density D
max
, then select all local maxima (points that
have the highest density in a circle of radius w) that have a density at least 0.3D
max
. These points
form the set of candidate orbit axes.
The next step is to nd arcs centered at these axes. I form a set of candidate orbits by considering
all pairings of orbit axes o and input images I
j
. I only accept candidate orbits that satisfy the three
constraints enumerated above, i.e., that the point of convergence is in the eld of view, the tilt is less
than 45

, and a sufcient number of points are visible in front of the orbit radius. I then evaluate
S
orbit
(o, I
j
) for each such suitable combination of orbit axis and image.
138
Figure 6.6: Panoramas detected in the Pantheon. The panorama detector found ve panoramas in
the Pantheon collection, shown in this image as blue circles (the two panoramas near the top of
the image are quite close to each other). For this scene, it is common for people to stand near the
circumference of the interior and take photos looking across the building in different directions.
I now have a set of orbits and a score for each orbit. To form a nal set of orbits, I select the
orbit with the highest score, remove it and all similar orbits from the set of candidates, then repeat,
until no more orbits with a score of at least 0.5 times the maximum score remain. Two orbits are
deemed similar if the area of intersection of the two circles dened by the orbits is at least 50% of
their average area. An example of detected orbits for the Pantheon collection is shown in Figure 6.5
and in the companion video. In this case, three orbits were detected, two around the outer facade
at different distances, and one around the altar in the interior. Note that the furthest detected orbit
arc is actually some distance behind the reconstructed cameras. The reason why this orbit does not
pass closer to the camera centers is the constraint that it should not be too similar to an existing orbit
(in this case, the inner orbit around the facade, which is the highest-scoring orbit). Furthermore,
this orbits still has a high camera score, as translating a viewpoint backwards from an image along
the viewing direction does not affect the angular deviation of rays nearly as much as translating the
viewpoint sideways.
When computing viewpoint scores for orbit samples, I use a default vertical eld of view of 50
degrees and a default screen resolution of 1024 768 (for the purposes of computing the eld of
view and resolution scores).
139
Figure 6.7: Canonical views detected in the Pantheon data set. Left: the rst nine detected canonical
views, using the algorithm of Simon et al.[130]. These views include the front facade, altar, oculus,
and several sculptures. Right: the canonical views shown in the interface as a row of thumbnails
at the bottom of the screen. Clicking on a thumbnail moves the user along an optimal path to the
corresponding view.
Panorama detection. A panorama consists of a set of images all taken close to a nodal point,
but possibly pointing in different directions. Similar to orbits, a good panorama has good views
available from a wide range of viewing directions. To nd panoramas in a scene, I rst consider
each image I
j
to be the center of a candidate panorama and compute a panoramas score S
pano
(I
j
)
for each candidate. S
pano
(I
j
) is computed as the sum of view scores S
view
for a range of viewing
directions around I
j
:
S
pano
(I
j
) =
25

=5

360

=0

S
view
(v(I
j
, , )) (6.11)
where v(I
j
, , ) is the viewpoint located at the optical center of image I
j
with viewing direction
given by angles (pan) and (tilt). Once each candidate panorama has been scored, I select a set
of nal panoramas from among this set. To do so, I select the top scoring candidate I

, remove all
images that have a non-zero reprojection score for some view v(I

, , ), then repeat this selection

process, until no remaining candidates score is above a threshold. The panoramas detected in the
Pantheon collection are shown in Figure 6.6, and as images in Figure 6.8.
140
Selecting canonical views. To compute the set of representative, canonical views, I use the scene
summarization algorithm of Simon et al.[130]. This algorithm seeks to capture the essence of a
scene through a small set of representative images that cover the most popular viewpoints. To select
a set of canonical views, the algorithm rst represents each image I
j
as a feature vector f
j
, where
f
j
has an entry for every 3D point p
k
in the scene, and where the kth entry of the vector, f
j
(k) = 1
if p
k
is visible in image I
j
, and 0 otherwise;

f
j
denotes a normalized version of f
j
. The scene
summarization algorithm then clusters the normalized feature incidence vectors

f
j
, then selects a
representative image for each cluster. The representative images are ordered by the size of their
respective clusters; the rst represents the most popular view, the second represents the second-
most popular view, and so on.
In addition to canonical views computed for the entire scene, I also choose a representative
image for each detected orbit and panorama. For a given set of images o associated with a given
control, the representative image is chosen as:
arg min
I
j
S

I
k
S,I
j
=I
k

f
j

f
k
, (6.12)
i.e., the image whose normalized feature vector is most similar to those of all other images in o.
5
6.5 Scene viewer and renderer
Once a scene has been reconstructed and augmented with scene-specic controls and canonical
views, it can be explored in the interactive scene viewer. The viewer situates the user in the scene,
exposes the derived controls to the user, and depicts the scene by continually selecting and warping
appropriate images as the user moves. This section describes the navigation interface and rendering
components of the viewer.
6.5.1 Navigation interface
As described in Section 6.4, the Pathnder viewer supports standard 3D translation, panning, and
zooming controls, as well as an orbit control, which rotates the camera about a xed 3D point or
5
For orbits, the denition of

f is slightly different; it only contains an entry for every point inside the orbit circle, as
only these points could possibly be salient to the object of interest.
141
Figure 6.8: Pathnder user interface. The scene viewer displays the currently photo in the main
view, and shows suggested controls in the thumbnail pane at the bottom of the screen. The pane is
currently showing detected panoramas. Arrows on the sides of the screen indicate which directions
the user can pan to nd more views.
axis. Typically, the orbital motion is constrained to a ring around an orbit axis, as many objects are
viewed from a single elevation, but our viewer also supports orbital motion on a sphere. When the
user is moving on a discovered scene-specic orbit, the orbit axis is automatically set to be the axis
discovered for that control. Alternatively, the user can manually specify an orbit point by clicking
on a 3D point in the scene. Once an orbit point is dened, the user can drag the mouse left and right
to orbit around a vertical axis, or up and down to move the viewpoint vertically (when 2D orbital
motion is enabled). The process is rapid and seamless: the user simply shift-clicks on a point in the
image (the closest 3D feature denes the orbit point), and the orbit begins as soon as the mouse is
moved. This orbit procedure is shown in the companion video for this chapter
6
for several scenes,
including the Statue of Liberty and the Venus de Milo. Figure 6.9 demonstrates how a user can orbit
around an object (in this case, a gurine of Mark Twain) to see it from different angles.
The user interface displays the derived scene-specic controls in a thumbnail pane at the bottom
of the screen. The user can choose between displaying pre-dened orbits, panoramas, canonical
6
http://phototour.cs.washington.edu/findingpaths/
142
Figure 6.9: Orbiting a home-made object movie of Mark Twain. This sequence shows screenshots
from a interaction session with a gurine of Mark Twain. The input images are from a video taken
of the gurine while rotating it around in front of the camera. The images show a user clicking and
dragging the mouse to rotate the object to the right (right image), left (middle image), and down
(left image).
images, and all controls, as shown in Figures 6.10, 6.7, and 6.8. Each control is depicted in the
pane with a thumbnail created from that controls representative image. The orbit and panorama
thumbnails are annotated with small arrow icons to show the type of control.
When the user clicks on a thumbnail, the system computes a path from the users current location
to the selected control (as described in the next section) and animates the virtual camera along that
path. If the user arrives at a panorama, left, right, up, and down arrows are drawn on the sides of the
screen indicating directions in which more images can be found, as shown in Figure 6.8. Similarly,
if the user arrives at an orbit control, left and right orbit arrows appear on the sides of the screen, as
shown in Figure 6.10. To determine if a particular arrow cue should be shown, the system computes
the viewpoint score for the view the user would see by moving in that direction a given amount. If
the score is above a threshold, the arrow is displayed.
6.5.2 Rendering
As the user moves through the scene, the system continually chooses an image to display based
on the reprojection score, S
proj
(I, v), which rates how well each database image I can be used to
render the current viewpoint v (Section 6.3). The image with the top score is selected as the next
image to be displayed. If no image has a large enough viewpoint score (i.e., S
view
(v) < ), no
image is displayed; only the point cloud is rendered.
If the input images are carefully captured and densely sample the space of all viewpoints (as
143
Figure 6.10: Detected orbits in the Statue of Liberty and Venus data sets. Left: two orbits were de-
tected for the Statue of Liberty data set, an inner orbit and an outer orbit. These orbits are displayed
to the user in the control panel at the bottom of the screen. Right: a single orbit was detected for the
Venus data set.
in Quicktime VR object movies and moviemaps), the rendering process is straightforwardsimply
display the photo corresponding to the desired viewpoint. In practice, however, casually acquired
image collections tend to be incomplete and irregularly sampled, with objects centered and oriented
differently in each image. Hence, it is necessary to warp each image to better match the desired
viewpoint. This is the function of the rendering engine.
To warp an image to match a desired viewpoint v, the image is projected onto a 3D projec-
tion surface in the scene, then into v. My system normally renders the scene using planar proxy
geometry, but can also render using a dense 3D model, if one is available.
Warping with proxy planes. When the system uses planar proxy geometry, it associates a plane
with each image and renders the image by projecting it onto the plane and back into the virtual view-
point. Planar proxies are also used in the Photo Tourism system, which ts planes to image points
for use in transitions. While these best-t planes work well for some scenes and navigation modes,
they can produce jerky motion in situations where the user is moving rapidly through a wide range
of views. This is especially true when orbiting; while the viewer is xated on a particular object,
the per-image best-t planes can stabilize different parts of the scene (including the background) in
different images, causing the object to jump around from frame to frame. A demonstration of this
144
o
D
D
S
P
o
S
o
D
O
D
S
P
o
S
O
Figure 6.11: Proxy planes should intersect the orbit point. Left: to warp an image from view S to
D, image S is projected onto the proxy plane P, which is then rendered into D. P passes through
the orbit point o, ensuring that o is rendered to the correct position in D. Right: P does not pass
though the orbit plane, causing o to be rendered to the wrong position.
effect can be found in the video on the project webpage [107].
My solution to this problem is simple but effective. Suppose the user is orbiting around an
orbit point o, indicating an object of interest, and suppose we wish to render the scene captured
by a source photo S into the destination viewpoint D. Consider a proxy-plane P in the scene.
I compute the warp by perspectively projecting S onto P, then back into D. As shown in Figure
6.11, making P intersect the orbit point o ensures that o projects to the correct location in D. Hence,
we can stabilize o in the rendered images (make o project to the same pixel (x, y) in all views) by
(1) choosing P to intersect o for each input view, and (2) orienting the rendered views so that the
viewing ray through (x, y) for each view passes through o. While any choice of P that passes
through o will sufce, choosing P to be parallel to the image plane of S results in well-behaved
warps (Figure 6.12). When an orbit axis, rather than a single point, is to be stabilized, we dene the
normal to P to be the projection of the viewing direction of S onto the plane orthogonal to the axis.
I call this form of image stabilization orbit stabilization.
Orbit stabilization performs a similar function to software anti-shake methods that reduce jitter
in video. However, it has the advantage of performing a globally-consistent stabilization, producing
the effect of rotation about a single center, and avoiding the drift problems that can occur with
frame-to-frame video stabilization methods. Note also that orbit stabilization does not require any
145
x
D
S
P
D
S
P
x x
x
Figure 6.12: Orientation of proxy planes. Left: if P is parallel to Ss image plane, a point x near the
orbit point will get mapped to a nearby point x

on P, causing a small error in D. Right: an oblique

choice of P will generally result in larger errors.
knowledge of scene geometry, although it does require known camera viewpoints and a reasonable
orbit point.
My system defaults to using best-t planar proxies, until an orbit point is selected, at which point
it switches to orbit stabilization. The user can also opt to use best-t planes even when an orbit point
is selected, which can produce better results if the scene is truly planar or nearly planar.
Warping with a 3D model. When a 3D scene model is available, it can be used in place of the
planar proxy to further improve rendering quality. Such models can be obtained, e.g., through laser
scanning [85], multi-view stereo [55], or manual modeling. To render an image with a 3D proxy, I
project the source image onto the proxy and then into the destination image, and place an additional
plane in back of the model to account for unmodeled geometry.
Using a 3D model for rendering usually results in a more realistic experience, but, as in the Photo
Tourism system, can also result in artifacts due to holes in the model or from projecting foreground
objects onto the geometry. In my experience, planar proxies tend to produce less objectionable
artifacts in these situations. The companion video shows a comparison of orbiting around the central
portal of the Notre Dame Cathedral using planar and 3D proxies. 3D proxies work well in this
example, and cause the 3D shape of the portal to be more pronounced. The artifacts resulting from
3D proxies are less noticeable here, as the facade of the Cathedral is roughly planar, hence projecting
foreground objects onto the proxy is less objectionable. On the other hand, for the Statue of Liberty
146
Figure 6.13: Screenshot from the Notre Dame collection. This screenshot from an interactive
Pathnder session exploring the Notre Dame collection shows multiple images blended together
as the user moves around the scene.
collection, these artifacts are more pronounced, and hence the 3D proxies work less well for that
data set.
Compositing the scene. The output image shown on the screen is rendered by rst drawing a
background layer consisting of the reconstructed point cloud drawn on top of a solid color, then ren-
dering the currently selected image, projected onto its proxy geometry. Rather than instantaneously
switching between images as new ones are selected for display, images are faded in and out over
time. The system maintains an alpha value for each image; whenever a new image is selected for
display, the alpha of the previous image decays to zero, and the alpha of the new image rises to
one, over a user-specied interval of time. When the user is moving on a planned path, the system
can look ahead on the path and fade images in early, so that each image reaches full opacity when
it becomes the optimal image. When the user moves on a free-form path, this prediction is more
difcult, and the system instead starts fading in an image at the moment it becomes optimal.
If the user is moving fairly quickly, multiple images can simultaneously have non-zero alphas.
I blend the images by rst normalizing all alphas to sum to one, then compositing the rendered
images in an off-screen buffer. This image layer is then composited onto the background layer. An
example screenshot showing multiple blended images is shown in Figure 6.13.
Once the photos are cached in memory, the Pathnder viewer runs at over 30 frames per second
147
with up to 1700 photos on a machine with a 3.4GHz Intel Xeon processor and an nVidia Quadro FX
4000 graphics card. A companion video showing the viewer being used to explore several scenes
can be found on the project website [107].
6.6 Path planning
As described in the previous section, the Pathnder viewer allows a user to select among the set of
detected controls and representative views using the thumbnail pane, and takes the user along an
automated path to a selected control. Unlike in Photo Tourism [137], which performs two-image
morphs, the Pathnder system can create transitions involving multiple images, making it much
more effective for moving along long, complex paths. These multi-image transitions are created
using a new path planning approach. The system attempts to nd smooth paths between two images
that are always close to good views, so that at any point on the path the user can be presented with
a high-quality rendering of the scene. An additional benet of constraining the path to pass near
photos in the database is that it will more closely emulate how a human would move through the
scene, as it will stay close to places where people actually stood and took photos. Such paths are
more likely to be physically plausible, e.g., to not pass through walls or other obstacles.
Path planning, often used in robotics, has also found application in computer graphics for com-
puting camera paths through 3D environments. For instance, Drucker and Zeltzer [38] use planning
to help create paths through a 3D scene which satisfy task-based objectives (such as focusing on a
specic object) and geometric constraints. In the realm of IBR, Kang et al.[76] analyze a sequence
of images to predict which views or portions of views can be synthesized. In my work, I extend
these ideas to use our view quality prediction score to plan good camera paths.
In order to nd the best path between two views given our database of existing images 1, sup-
pose we have a viewpoint cost function Cost
I
(v) (to be dened shortly) over the space of possible
viewpoints, where Cost
I
is low for viewpoints close to existing image samples, and large for more
distant views. The optimal path between two viewpoints can then be dened as the lowest-cost path
(geodesic) connecting them.
The dimension of the viewpoint space, however, is relatively high (ve degrees of freedom for
camera pose, and six if zoom is included), and therefore this continuous approach is computationally
148
expensive. I instead nd a discrete solution by rst computing an optimal piecewise linear path
through the set of existing cameras, then smoothing this path. This discrete problem can be posed
as nding a shortest path in a transition graph G
T
whose vertices are the camera samples 1 (G
T
is
a weighted version of the image connectivity graph presented in Chapter 3).
G
T
contains a weighted edge between every pair of images I
j
and I
k
that see common 3D
points. The weight w
jk
on an edge (I
j
, I
k
) is the predicted cost of a transition, computed based on
how well the transition between the two images can be rendered. We denote this transition cost as

jk
, and dene
jk
to be the integral of the viewpoint cost function over a straight-line path
jk
(t)
between I
j
and I
k
.
7
Because edge (I
j
, I
k
) represents a two-image transition, when computing
jk
,
I restrict the rendering process to consider only I
j
and I
k
when generating in-between views along
the path
jk
(t); this restricted viewpoint cost function is denoted Cost
jk
(v). Thus,

jk
=
_
1
0
Cost
jk
(
jk
(t)) dt. (6.13)
I dene Cost
jk
by rst considering the cost of rendering a new viewpoint with one of the images
samples, I
j
. In Section 6.3 I dened a reprojection score S
proj
(I
j
, v) for estimating how well an
image I
j
can be used to synthesize a new view v. I now turn this scoring function into a cost
function:
Cost
j
(v) = e
(1S
proj
(I
j
,v))
1. (6.14)
This function evaluates to 0 when S
proj
(I
j
, v) = 1, and to e

1 when S
proj
= 0 (I use a value
= 8). I now dene the two-view cost function Cost
jk
over the path
jk
(t) as the weighted sum
of the single viewpoint cost function:
Cost
jk
(
jk
(t)) = (1 t) Cost
j
(
jk
(t)) +t Cost
k
(
jk
(t)). (6.15)
I approximate the integral in Eq. 6.13 by computing the sum of Cost
jk
at 30 samples along
jk
(t).
If edges are weighted using the transition cost alone, shortest paths in the graph may not corre-
spond to smooth paths through viewpoint space; indeed, I have observed that such paths can zigzag
7
A straight-line path is obtained by linearly interpolating the camera position and quaternions representing each cam-
era orientation.
149
Figure 6.14: Using path planning to compute a camera transition. A transition from an image
outside the Pantheon (green) to an image inside (red) computed using my path planning algorithm.
The blue cameras are the intermediate nodes visited on the transition graph, and the blue line is
the linearly interpolated path. The black curve shows the path resulting from smoothing this initial
discrete path, and the red lines indicate the viewing directions at samples along this path.
in both position and orientation in a disorienting way. To penalize such paths, I add a smoothness
cost
jk
to each edge weight w
jk
. This cost is simply the length of the edge in viewpoint space,
which I compute as a weighted combination of the difference in position and orientation between
cameras C
j
and C
k
:

jk
= |c
j
c
k
| + angle(v(C
j
), v(C
k
)), (6.16)
where c is the 3D position of camera C and v(C) is its viewing direction. I have found a value
= 3.0 to work well in practice.
The nal weight of an edge (I
j
, I
k
) is the weighted sumof the transition cost
jk
and the smooth-
ness cost
jk
:
w
jk
=
jk
+
jk
. (6.17)
For my experiments, I use a value of = 400.
150
Generating smooth paths. To generate a path between two images I
start
and I
end
, I rst use
Djikstras algorithm [35] to compute a shortest path

between I
start
and I
end
in G
T
.

can be
interpreted as a piecewise linear physical path through viewpoint space; this path is next smoothed
to produce a more continuous path for animating the camera.
To smooth

, I rst uniformly sample

to produce a sequence of viewpoint samples v

0
i
, i = 1
to n (I use n = 100 samples in my implementation). I then repeatedly average each sample with its
two neighbors, and with its original position v
0
i
(in order to keep the sample close to the initial path

):
v
t+1
i
=
1
1 +
_
0.5
_
v
t
i1
+v
t
i+1
_
+v
0
i
_
. (6.18)
I apply 50 iterations of smoothing.
8
The parameter controls how closely the nal path matches

, versus how smooth the path is. In my implementation I set = 0.02, which produces nice,
smooth paths which still generally stay close enough to the images along

to produce good views.

An example of a path computed between two images in the Pantheon collection is shown in Figure
6.14.
The nal path
nal
between I
start
and I
end
is formed by linearly interpolating the nal view-
point samples v
i
. To animate the camera, I sample
nal
non-uniformly, easing in and out of the
transition. The user need not be situated exactly at the position of I
start
when the transition is started;
I adjust the initial path

to begin at the users current position before smoothing. If the user is not
visiting any image, however (i.e., the user is at a viewpoint v where S
view
(v) < ), then my system
currently uses a simple linear path to move to I
end
. This limitation could be addressed by adding a
virtual node to G
T
corresponding to the users current viewpoint, adding edges between this node
and nearby images, then weighting these edges with an appropriate cost function.
6.7 Appearance stabilization
Unstructured photo sets can exhibit a wide range of lighting and appearance variation, including
night and day, sunny and cloudy days, and photos taken with different exposures. By default,
Pathnder displays photos based only on the users current viewpoint, which can result in large
changes in scene appearance as the user moves. These large, random variations can be useful in
8
This smoothing operation is equivalent to applying a Gaussian lter to the path.
151
(a) (b) (c) (d) (e) (f)
Figure 6.15: Computing image similarity. Image (a) with weight map (b); higher weights are darker.
To compare (a) to another image (c), I rst downsample (a) to a height of 64 pixels (d) and resize
the second image to match the scale of the rst. I then geometrically warp the second image to be
in alignment with the rst (e), and apply color compensation to the rst image (f). The distance
between the images is then computed as the weighted L
2
distance between the RGB pixel values of
(e) and (f).
getting a sense of the variation in appearance space of a scene, but they can also be visually distract-
ing. To reduce this appearance variation, the user can enable a visual similarity mode, which limits
transitions to visually similar photos, and a color compensation feature, which hides appearance
changes by modifying the color balance of new images to better match that of previous images. In
addition, our system allows the photo collection to be separated into classes, such as day and night,
and allows the user the option of restricting selected photos to a particular class.
This section describes these features in more detail. I rst describe a metric used to compute
photo similarity, then describe how this metric is incorporated into the viewer and how color com-
pensation is done. Finally, I describe how an object can be browsed in different appearance states.
6.7.1 Computing image distances
To reduce the amount of appearance variation that occurs while viewing a scene, I rst need a
way to measure the visual distance between two images. To compute this distance, I register the
images geometrically and photometrically, then compute the L
2
distance between corresponding
pixel values, weighted by condence in the registration. These steps are summarized in Figure 6.15.
Geometric alignment. To compute the distance between images I
j
and I
k
, I rst downsample I
j
to a resolution of 64 64 pixels, and downsample I
k
to approximately the same sampling rate with
respect to the scene, resulting in low-resolution images I

j
and I

k
.
152
Next, I warp I

k
into geometric alignment with I

j
. If the complete scene geometry is known, it
can be used produce the warped version of I

k
, but since I do not assume geometry is available, I
instead use the sparse data interpolation method of thin-plate splines (TPS) [12] to model a non-rigid
2D warp between images.
9
Given a set of 2D image correspondences (x
i
, y
i
) (x

i
, y

i
), i = 1 to
k, a thin plate spline models the image warp D(x, y) = (D
x
(x, y), D
y
(x, y)) separately for each
dimension using a combination of an afne transformation and radial basis functions:
D
x
(x, y) = a
1
+a
x
x +a
y
y +
k

i=1
c
i
([[(x
i
, y
i
) (x, y)[[) (6.19)
D
y
(x, y) = b
1
+b
x
x +b
y
y +
k

i=1
d
i
([[(x
i
, y
i
) (x, y)[[) (6.20)
where the kernel function (r) = r
2
log(r) and the coefcients a, b, c, and d are the parameters of
the model. These parameters can be solved for in closed form.
To compute the warp, I project all 3D points visible to I
j
into both I

j
and I

k
to form a set of 2D
basis points, and compute the corresponding TPS deformation D mapping I

j
onto I

k
. Computing
the TPS parameters involves solving a large, dense linear system Ax = b. In my implementation,
I use an LU decomposition to solve the system; because Aonly depends on the projections in I

j
, I
pre-factorize A into L and U in order to quickly solve for the TPS parameters each time an image
is registered with I
j
.
Given the deformation D, for each pixel location x = (x, y) of I

j
, I compute the corresponding
pixel location D(x) of I

k
. If D(x) lies inside I

k
, I

k
is sampled at D(x) using bilinear interpolation.
This results in a sequence of pairs of RGB samples:
[I

j
(x
1
), I

k
(D(x
1
))], [I

j
(x
2
), I

k
(D(x
2
))], . . . , [I

j
(x
n
), I

k
(D(x
n
))]
Figure 6.15 (e) shows an example of one image warped into another using this technique.
The TPS deformation D will not necessarily extrapolate well far away from the basis points;
there is more condence in the deformation near known 3D point projections. To make the image
9
Planar proxies could also be used to warp the images into alignment, but I have found that, while planar proxies
often create reasonable visual results, they are not always accurate enough for the purpose of registering images for
computing image differences.
153
comparison more robust to misregistration, I precompute a spatial condence map W
j
for each
image I
j
. W
j
is created by centering a 2D Gaussian at the projection of each 3D point observed
by I
j
, with standard deviation proportional to the scale of the SIFT feature corresponding to that
point, and with height
1
2
. The Gaussians are then summed, sampled at each pixel, and clamped to
the range [0, 1]. Figure 6.15 (b) shows an example condence map.
When registering image I

k
to image I

j
, I also create a condence map for the registered result.
To do so, I downsample the weight maps W
j
and W
k
in the same way as the images, producing
low-resolution weight maps W

j
and W

k
, then create a weight for each of the RGB sample pairs
[I

j
(x), I

k
(D(x))] by taking minimum of the two condence values corresponding to these samples:
w
i
= minW

j
(p
i
), W

k
(D(p
i
)) (6.21)
giving a sequence of weights, w
1
, w
2
, . . . , w
n
.
Photometric alignment. After applying a spatial warp to I

k
, I next align the color spaces of the
two images. I use a simple gain and offset model to warp each RGB color channel of I
k
to match
that of I
j
:
M
kj
=
_

_
s
r
0 0 t
r
0 s
g
0 t
g
0 0 s
b
t
b
_

_
This afne transform can compensate linear changes in brightness, due, for instance, to changes in
exposure time. To compute the gain and offset parameters s and t for each color channel, I t a
line to the pairs of color samples; I use RANSAC [45] during the tting to achieve robustness to
bad samples due to misaligned or saturated pixels [45]. Figures 6.15 (e) and 6.15 (f) show a pair of
geometrically and photometrically warped images.
Finally, I compute the distance measure
d
(I
j
, I
k
) as the weighted average of color-shifted RGB
samples:
d
(I
j
, I
k
) =
1

w
k
n

i=1
w
i
_
_
I

j
(x
i
) M
kj
I

k
(D(x
i
))
_
_
(6.22)
using RGB values in the range [0, 255].
154
Because it is difcult to reliably register photos with wide baselines, I only compute image
distances between pairs of photos that are relatively close to each other. In the Pathnder viewer,
it is still possible to move a large distance when similarity mode is enabled, via a sequence of
transitions between nearby, similar images.
6.7.2 Browsing with similarity
The Pathnder system reads the pre-computed similarity scores at startup. When the visual similar-
ity mode is enabled, these scores are used to prune the set of possible image transitions and to favor
transitions between images that are more similar. This is implemented by multiplying the reprojec-
tion score S
proj
(I, v) with a similarity factor S
sim
(I, I
curr
), where I
curr
is the currently displayed
image. To compute S
sim
(I, I
curr
), I remap the interval [d
min
, d
max
] to [0, 1] and clamp:
S
sim
(I, I
curr
) = 1 clamp
_
d
(I, I
curr
) d
min
d
max
d
min
, 0, 1
_
. (6.23)
I use values of d
min
= 12 and d
max
= 30. Enabling similarity mode results in the selection of a
sparser set of photos, so though their visual appearance is much more stable, the motion can be less
continuous.
6.7.3 Color compensation
The viewer can also use the pairwise RGB gain and offset parameters, estimated while computing
similarity scores, to disguise changes in appearance by adjusting the color balance of a new image
to match that of the previous image. When color compensation is enabled, the viewer maintains a
3x4 color compensation matrix M
j
for each image I
j
, which is applied when an image is rendered.
During a transition from an image I
j
to an image I
k
, M
k
is set to M
kj
, pre-multipled by the color
compensation already in effect for I
j
,
M
k
= M
j
M
kj
.
Examples of color corrected images are shown in Figure 6.16. To reduce problems with accumulated
drift over time and eventually return images to their true color balance, the matrices M
j
fade back
155
Figure 6.16: Similarity mode and color compensation. The rst row shows a sequence of images,
going from left to right, from an object movie of the Trevi Fountain resulting from orbiting the site.
The second row shows the result of orbiting through the same path with similarity mode turned on.
Note that a different set of images with more similar lighting is selected. The third row shows the
same images as in the second row, but with color compensation turned on. The color balance of
each image now better matches that of the rst.
to the identity transform [I[0] over time.
6.7.4 Viewing different appearance states
As with QuickTime VR object movies, my system allows an object to be viewed in different ap-
pearance states, such as day and night. This feature requires the photos to be classied into sets
corresponding to each state; once the photos are classied and a user selects a certain state, the
system will only display photos from that state. At a given viewpoint v, the user can toggle to any
state for which a photo I
j
with non-zero reprojection score S
proj
(I, v) exists.
I found that for two particular classes of photos, night and day, could be semi-automatically
classied using the observation that many 3D points (corresponding to SIFT features in the original
images) are highly correlated with either daytime or nighttime images; some features appear only
in daytime images, and some only in nighttime images. By hand-labeling a small number daytime
and seven nighttime photos, labels can be automatically propagated to all other photos.
156
Figure 6.17: Switching between day and night in the Trevi Fountain. Left: a daytime view of the
Trevi Fountain. Right: the result of switching to a nighttime view; the same viewpoint is shown at
a different time of day.
I In particular, the labeling algorithm iteratively updates a set of (continuous) image labels U(I)
and point labels V (I) [1, 1], where -1 corresponds to a nighttime image and 1 to a daytime
image. The image labels are rst initialized to U(I) = 1 for the manually labelled daytime images,
U(I) = 1 for the nighttime images, and U(I) = 0 for other images, and the point labels were
initialized to V (p) = 0. Next, the point and image labels are updated using the following update
equations:
V (p) =

IImgs(p)
U(I)

IImgs(p)
[U(I)[
, U(I) =
1
[Points(I)[

pPoints(I)
V (I),
where Points(I) is the set of points visible in image I, and Imgs(p) is the set of images in which
point p is visible. In other words, points that are seen in mostly night (resp. day) images are
labeled as night (resp. day) points, and vice versa. For the cases I tried (the Trevi and Notre
Dame sets described in the next section), the update steps converged after about ve iterations, and
cleanly separated the images into daytime (U(I) > 0) and nighttime (U(I) < 0) sets.
6.8 Results
I have applied the Pathnder system to several large collections of images downloaded from Flickr,
as well as a collection of images taken from a hand-held camera. Screenshots of the interactive
viewer are shown in gures throughout this chapter (Figures 6.5, 6.10, 6.8, 6.9, and 6.13 and the
gures in this section); however, the results are best viewed in the video on the project web site [107].
Two of these scenes consist of dominant objects and provide an object movie experience: the Statue
157
of Liberty, created from 388 photos, and the Venus de Milo, created from 461 images. The system
detected two orbits for the Statue of Liberty, and one orbit for the Venus de Milo; these orbits are
shown in Figure 6.10. The reconstruction of the Notre Dame Cathedral (created from 597 photos)
has a wide distribution of camera viewpoints on the square in front of the Cathedral, and is therefore
well suited for free-form 6-DOF navigation; a screenshot of the viewer with this data set loaded is
shown in Figure 6.13. This is a case where automatic orbit detection is less useful, as a good orbit
can be produced from almost anywhere on the square, as shown in the video. The reconstruction
of the Trevi Fountain (1771 photos) contains large numbers of both day and nighttime images,
making this a good candidate for evaluating both appearance stabilization and state-based modes.
Figure 6.17 shows an example of switching between day and night mode.
I also demonstrate how the system makes it easy to create an object movie experience by manu-
ally rotating a hand-held object (a Mark Twain gurine) in front of a camera. In this case, the user
manually specied a sphere of orbits, as the current implementation does not support spherical orbit
detection; screenshots are shown in Figure 6.9.
Finally, I demonstrate the system with a collection of photos of the Pantheon (602 images), a
complex scene consisting of both interior and exterior views. For this scene, Pathnder detected
three orbits (Figure 6.5), several panoramas (Figures 6.6 and 6.8), and a number of canonical im-
ages (Figure 6.7), including photos of the front facade, the altar, the oculus, and several sculptures
inside the building. The accompanying video shows sample interactions with each of these types of
controls, and demonstrates the results of my path planning approach.
I also use this reconstruction to demonstrate another application of the Pathnder system: cre-
ating a 3D slideshow of a personal photo collection. Suppose that you visit the Pantheon and take
your own set of photos, showing friends and family in a few discrete locations (e.g., the front facade,
the altar, looking up at the oculus). Structure from motion techniques are not likely to be able to reg-
ister the photos together due to insufcient overlap, unless they are captured much more frequently
than is customary with personal collections. However, if we combine them with all the other photos
of the Pantheon on the Internet, we can register our personal photos with that collection, and plan
paths between our own photos to create a 3D slideshow (in a sense, retracing our steps through the
Pantheon). The companion video shows such a slideshow created from a collection of four personal
photos added to the 602-image reconstruction. Figure 6.18 shows several photos in this personal
158
Figure 6.18: Personal photo tour of the Pantheon. In this sequence, a user has added their own
personal collection to the Pantheon data set. This allows the user to create a 3D slideshow through
their own photos, using the communitys photos to ll in the gaps (please see the video on the project
webpage for an animated version of this tour).
collection in context in the Pathnder viewer.
6.9 Discussion
I have successfully used my approach to create uid IBR experiences with scene-specic controls
from unstructured community photo collections. I believe that these techniques represent an impor-
tant step towards leveraging the massive amounts of imagery available both online and in personal
photo collections in order to create compelling 3D experiences of our world. Part of the power
of this approach is the ability to learn controls from and create renderings of the world from large
photo collections. As these techniques are applied to larger and larger collections, I believe that the
experiences generated will continue to improve.
However, my approach also has several limitations in both the navigation and rendering compo-
nents. My geometric model for orbits is a circle, which ts many real-world scenes. On the other
hand, many paths around objects are ellipses, lines and polygons, or more free-form shapes. For
instance, consider the reconstructions of the Pisa Duomo and Stonehenge discussed in Chapter 4
and reproduced above in Figure 6.19. The paths people take around these objects follow walkways
or (in the case of Stonehenge) fences, resulting in more complex, non-circular paths. In the future,
it would be interesting to explore the detection of more general types of paths in a scene, perhaps
by unifying the path planning algorithm with the orbit and panorama detection algorithms. An ad-
ditional challenge is to devise better rendering algorithms for these more general paths, as orbit
stabilization will not always be applicable.
159
(a)
(b)
Figure 6.19: Free-form paths around the Pisa Duomo and Stonehenge. Not all paths through scenes
are circular. (a) Photos taken of the Pisa Duomo and its surroundings tend to follow walkways
around the building. (b) Photos of Stonehenge follow an arc going about two-thirds of the way
around the prehistoric monument, but then curve inwards to follow a straight section of path for the
remainder of the circle.
Another interesting question brought up by this point is whether the paths a user wants to take
through a scene are the same as the ones people actually take when physically at the scene. The
paths discovered by my system reect the latter, which results in controls that are arguably useful in
two senses: (1) they show you what you would presumably see if you were there and (2) inasmuch
as people take photos from optimal viewpoints and of interesting views, the derived controls are
likely to be interesting as well. However, people are also constrained by physics, walls, and other
physical barriers, which preclude viewpoints that might, in some sense, be better. For instance, the
fences around Stonehenge prevent people from getting close to and climbing around the stones, but
it would certainly be useful to have photos taken from such viewpoints (in fact, these views might
be much more useful in a virtual setting, where there is no possibility of damage to the stones). One
possible way around this problem might be to combine the controls learned from the masses with
more carefully created paths captured by experts with special access or equipment, or to have the
160
masses vote on interesting paths through the scene with a distributed 3D drawing interface.
An expert creating a specialized path through a scene is an example of the more general concept
of authoring. Currently, the system discovers sets of controls and computes paths between views
completely automatically and without user input (the user can specify a set of endpoints to paths,
however, as in the 3Dslideshowshown in Section 6.8). In some cases, a user might want to manually
specify paths, either to correct a mistake made by the automatic algorithm, highlight additional
interesting paths, or create a personalized tour through the scene. For instance, one could envision
a kind of magnetic lasso tool [96], where a user draws a path that automatically snaps to nearby
photos so as to improve the viewpoint score of the specied path.
Appearance compensation is also still a signicant open problem. My color compensation
method works well for images that are fairly similar in overall appearance, so it goes hand in hand
with the similarity-based selection mode. However, because the method only models an afne trans-
formation per color channel, compensating two very different images (e.g., sunny and cloudy) is not
possible, limiting the number of possible transitions between images. Developing a more exible
appearance compensation model would help avoid these problems. It also would be interesting to
explore more sophisticated image models that detect foreground objects, such as people, and back-
ground elements such as the sky. These elements could then be treated specially during rendering:
foreground objects could be removed during transitions, or popped up onto their own proxy planes,
and the sky could transition in a realistic way, with clouds rolling in or out.
For larger and larger scenes, memory will become a signicant issue. Currently, the system
caches all images when the program is loaded: this not only introduces a signicant startup time,
but will not scale to collections of tens or hundreds of thousands of images. Handling truly massive
collections will require more efcient caching schemes, such as those used in Seadragon [127] and
Photosynth [109].
How will the controls themselves scale to much larger scenes? For a city-sized scene, will
the user want the same set of controls as for a small-scale scene? I imagine that the toolset for
moving through the scene would probably consist of controls at different levels of detail, depending
on whether the user wants to get an overview of the entire city or a particular momument. Paths
between different parts of a city could be computed in a variety of ways as well, depending on the
users intent. A Google-Earth style transition (zooming out to city-level, then zooming back in to
161
the destination) would provide a broad context of the scene, while a taxi-style transition wending
through the city streets would show how a person might actually move between two parts of the city.
Evaluation. Finally, evaluation is a very important area for further study. How effective are the
derived scene-specic controls, and the Pathnder user interface as a whole, at enabling users to nav-
igate through and understand a scene? Part of the challenge in evaluating 3D navigation techniques
is deciding what questions to askwhat is it that we want to evaluate? Gauging the effectiveness
of an interface presupposes a certain task that users are trying to accomplish, as well as a certain
metric for success, such as speed or accuracy. For instance, in Ware and Osbournes study compar-
ing different navigation metaphors [157], the task is to locate several details distributed throughout
the scene (and afterwards make a movie showing where each detail is), and the evaluation metric
was how satised the user was with each metaphor. Other work has considered the efciency of
completing tasks such as moving the viewpoint to a specic object [145] (somewhat analogous to
the 2D desktop task of moving a mouse cursor to a destination).
In practice, however, it can be hard to predict what task a user will wants to accomplish in
a scene, or the task itself may be ill-dened. In the context of this chapter, which has focused
mainly on famous, real-world scenes, the most natural task is perhaps virtual tourismwalking
around a scene doing what one might do if physically visiting a famous place for the rst time,
such as looking at interesting objects and learning interesting facts about the location. These kind
of tasks seem less straightforward to evaluate, as they are more about entertainment and education,
or, perhaps, simply about the experience of being at an impressive or important place. For these
kinds of tasks, a more informal type of study evaluating how engaged users are in the interface, and
how long they remained interested, might be more appropriate. It would also be interesting to study
how people move through and interact with real tourist environments in order to create improved
interfaces and experiences. The work presented here uses an artifact of those interactionspeoples
photographsto reverse-engineer patterns of movement and nd interesting objects, but studying
behavior more directly could also be fruitful.
In addition to virtual tourism, there are also other scenarios that would be interesting to evaluate.
One example is real estate, where an owner might want to create a virtual tour of their home for po-
tential buyers to explore. From a buyers point of view, the goals of exploring a home might be more
162
specic than with tourist sites, and could include seeing what individual rooms look like, how big
they are, and how they are spatially arranged. One way to evaluate a virtual home interface would
be to measure how well (i.e., how accurately, and how quickly) a user gains spatial knowledge of a
home when using the interface. This could be tested in a study in which users explore the home vir-
tually, then try to sketch out a oorplan. This approach was used by Darken and Seibert to evaluate
how well different navigational aids (e.g., an overhead map or a grid) improve waynding ability
in large virtual environments [27]. Other measures of spatial knowledge have also been explored.
For instance, Ruddle et al.[121] use three different metrics in the context of navigating virtual build-
ings: route-nding ability, relative distance between landmarks, and directional estimates, i.e., how
accurately users could point from one location to another after exploring a scene.
Individual components of the Pathnder system could also be evaluated. For instance, the effec-
tiveness of the scene-specic controls could be evaluated in comparison to other types of controls
(e.g., different styles of free-viewpoint controls). The trajectories resulting from the path planning
algorithm could also be compared to other types of paths, such as linear interpolation, in terms of
scene and motion comprehension or visual appeal.
163
Chapter 7
CONCLUSION
This thesis has presented my work on taking large, unstructured photo collections and creating
immersive 3D experiences of places, using combinations of new computer vision, graphics, and
interaction techniques. This work makes the following specic contributions to computer vision
and computer graphics:
Computer vision:
A structure from motion (SfM) pipeline, described in Chapter 3, that combines many
existing techniques (SIFT [90], RANSAC [45], sparse bundle adjustment [89]), into a
robust system. With this system, I have demonstrated, for the rst time, that SfM can
be successfully applied to the large, unstructured, highly diverse collections of images
found with Internet search.
An approach for handling the scale of Internet collections. In Chapter 4, I described
skeletal sets, an approach for sifting through a large image collection to nd the subset
that is critical for reconstruction. This technique can signicantly reduce the number of
images that need to be processed with the heavy machinery of SfM, reducing the SfM
processing time by an order of magnitude for large collections, while attaining provable
bounds on the loss in accuracy of reconstruction.
Computer graphics and interactive techniques:
A new 3D photo browser. In Chapter 5, I presented Photo Tourism, a photo browser
that takes photo collections reconstructed with SfM and places the user in the 3D scene
among the reconstructed photos. Photo Tourismprovides newgeometric photo browsing
controls that enable actions such as zooming in and out of a photo, moving to photos
taken to the left or right of a given image, nding good photos of a selected object, and
164
viewing stabilized slideshows. Photo Tourism also contains a powerful annotation tool,
which can be used to label large numbers of photos at a time.
A new 3D navigation system that automatically derives good navigation controls from
an input photo collection. Chapter 6 described Pathnder, a scene browser that ana-
lyzes the distribution of photos in a collection, detects interesting orbits, panoramas,
and canonical views, and synthesizes optimal paths between different controls using
a new path planning algorithm. Pathnder tailors the navigation controls to different
scenes in order to make exploring a particular scene more intuitive and automatic.
New rendering techniques for depicting scenes from unstructured photo collec-
tions, including new ways to render transitions between images and to render scenes
in a non-photorealistic style (Chapter 5), and an approach for selecting and display-
ing images automatically based on a users current viewpoint and navigation controls
(Chapter 6).
7.1 Future work
In this thesis, I have presented work that demonstrates the tremendous potential of using massive
photo collections to create new ways to visualize the world. However, this work is just a rst step
towards a goal of creating a virtual version of our world that is as visually rich and as extensive
as the real thing, putting a vivid recreation of the world at our ngertips. I envision a system that
stores all of the images of places ever taken and can display or synthesize a photo-realistic view
from nearly anywhere in the world, at any time, and under any possible weather condition. In this
section, I describe several specic projects that reach toward this ambitious goal.
Reconstructing Rome. Scale is still a signicant problem in structure from motion problems. To
date, the largest collections I have reconstructed consist of about 10,000 photos. However, for many
large-scale scenes, we can easily nd collections of photos that are two or three orders of magnitude
larger. For instance, searching for Rome on Flickr returns over 1.5 million photos. If we add the
Italian name for Rome, Roma, we get an addition half million photos. The same search in Picasa
Web Albums results in nearly 18 millions photos. I have downloaded just over a million of these,
and matched a subset of about 20,000 of the downloaded images. The image connectivity graph for
165
Figure 7.1: Image connectivity graph for 20,000 images of Rome from Flickr. The graph is made up
of several large (and many smaller) connected components corresponding to individual sites, some
of which are labelled.
this subset is shown in Figure 7.1. The goal is to reconstruct as much of the city of Rome as possible
from these photosto create a modern-day complement to the model of the ancient city created by
the Rome Reborn project [62].
How can we reconstruct one million photos? Further, can we reconstruct all of them in a single
day? This is a very challenging problem that will require new, more efcient algorithms, but also
intelligent ways of parallelizing the reconstruction effort.
Generally speaking, there are two parts of the reconstruction pipeline that are particularly time-
consuming: image matching and structure from motion. In the current image matching algorithm,
every pair of images is compared; for one million photos, we need to consider just under half a
trillion image pairs. At 0.1 second per pair, this process would take about a millenium and a half to
complete on a single machine. The matching is easily parallelizable, but even with 1,000 machines,
it would take more than 1.5 years (about 578 days) to complete. Linear-time algorithms based on a
166
bag of words approach [133, 104, 24] are much more scalable, but may be less accurate than brute
force matching. These approaches might be fruitfully combined together in stages. For instance,
one could rst run a linear-time algorithm to nd likely pairs of matching images, followed by a
brute force verier on those pairs, followed by a stage to compress the matching information to
remove redundancy, and nally another brute force step to nd any missed matches.
For the structure from motion stage, the skeletal sets algorithm presented in Chapter 4 (com-
bined with parallelization of the pairwise reconstructions) may be efcient enough to reconstruct
each connected component of the match graph. If not, however, one way to make SfM even faster
would be to parallelize the reconstruction of the skeletal graph by recursively breaking it apart,
reconstructing individual pieces in parallel, and merging.
Robustness is another challenge with this large reconstruction effort. As noted in Chapter 3, the
SfM pipeline has several failure modes that can result in incorrect results. For individual models,
these failures can sometimes be xed by manually tweaking parameters or by choosing a different
initial image pair. However, hundreds of reconstructions may be present in the million images of
Rome, and it would be tedious to check each one. Thus, improving the success rate of SfM is
another key problem.
Capturing appearance and dynamic phenomena. In this thesis I have focused mainly on re-
covering camera and sparse scene geometry from photo collections and providing corresponding
geometric controls. However, the appearance of a scene can change dramatically over time, from
day to night, over the course of the year, and due to changes in the scene itself. In the Pathnder sys-
tem, I demonstrated techniques for normalizing appearance in order to maintain consistency across
views, and for partitioning photos into broad categories, such as day and night. Even better would
be controls for directly exploring the space of the appearance of a scene, in the same way that geo-
metric controls let the user explore the space of viewpoints. Imagine, for instance, having a handle
on the sun, being able to drag it across the sky, and having the appearance of the scene automatically
adjust to reect the selected time of day and year.
One approach to this problem would be to index each photo by viewpoint and time of capture,
and display the photo that best matches the users current viewpoint and selected time. However, the
photo collections I have reconstructed are not yet dense enoughin both viewpoint and timefor
167
this simple lookup scheme to work well. Moreover, the time and date information that is stamped
on most digital photos can be unreliable, as photographers can easily forget to set their cameras
clock or update the time zone when they travel.
If the time and date information could be reliably determined, another approach would be to
learn an entire appearance model from scratch given the images. This would involve computing a
full 3D model for the scene, learning material parameters (or bi-directional reectance distribution
functions (BRDFs)) for every surface, determining how the scene is lit, and using these recovered
parameters to render the scene from new views and at new times. This is a challenging data-tting
problem, however, and would involve solving a complex, highly non-linear objective function; nev-
ertheless, this would be a very interesting approach to try. A somewhat less ambitious approach
would be to recover parameters of a simplied model from the data, using, e.g., PCA to learn a
linear basis for the scene appearance.
Of course, the appearance of a scene changes for other reasons besides variations in illumination.
At longer time scales, buildings are built and demolished
1
and at shorter time scales, people move,
water ows, and trees sway. Can we also capture and display these kinds of dynamic effects? For
shorter time scales, it would be interesting to incorporate video of a scene into a reconstruction, and
attempt to resynthesize dynamic elements from the video in other views.
Video. The dense temporal sampling provided by video is not only helpful in capturing dynamic
effects, but could also be useful in many other ways as well. For instance, video can record the
paths people take through a scene much more densely, and thus it would be interesting to apply the
analyses in Chapter 6 to image collections with intermingled videos. At an even more basic level, it
would be interesting to characterize how (and why) people tend to capture video of a scene or event,
and how this differs from how people capture photos. Audio often comes along for free with video;
adding audio to scenes would make for richer experiences, and could provide an authoring interface
for creating video tours. An interesting interface for browsing such video is presented in [117].
1
Usually in that order; this constraint is being used by the 4D Cities project to nd a temporal ordering of historical
photos of Atlanta [125].
168
Interiors. I have mainly explored reconstructing and exploring outdoor (usually architectural)
scenes. While indoor scenes are not necessarily more challenging to reconstruct (see, for instance,
the St. Peters and Pantheon reconstructions in Chapter 4), they do seem to be more challenging to
visualize through photo collections. One reason for this is that, while the exterior front of a building
can often be reasonably approximated with a plane, the inside of, say, a home, is more complex.
Typically, several walls (as well as oor and ceiling) are visible at any given time, and there tend
to be more objects (e.g., furniture) inside than outside. Hence, planar projection surfaces do not
work as well inside (the Pantheon is an exception, as the interior is approximately a simple cylinder
topped by a dome). To create good renderings of interiors, it may be necessary to use more complex
projection surfaces consisting of multiple planes or triangulated meshes.
Building a new photo-sharing community. The work presented in this thesis makes use of the
photos of many different photographers taking photos independently and without the intention of
collectively reconstructing 3D geometry or creating immersive experiences. What if people knew
that they were contributing photos towards a larger reconstruction effort, and could see the results
of that effort as they unfolded? It would be interesting to create a website where people could
upload their photos, as in Flickr, but where uploaded photos are automatically registered with an
existing set of reconstructions. The current state of this world reconstruction could be viewed at any
time, and the boundariesthe parts where new photos are neededcould be highlighted. Users
could also see their own photos, as well as the parts of the world to which they have contributed.
This distributed reconstruction effort could also be turned into a game in which points or other
incentives are offered for adding onto the reconstruction, or incorporated into other fun activities,
such as virtual treasure hunts. This type of online reconstruction procedure poses certain algorithmic
challenges. For instance, given an image collection and a corresponding skeletal graph, is it possible
to update the skeletal graph in an online fashion as new images come in? On the other hand, having
users more involved may simplify other problems. Image matching, for instance, may be much
easier if users interactively provide hints as to their location, e.g., by clicking or tapping on a map,
thus constraining candidate matching images to nearby photos.
Creating a website that collects and registers the worlds photos could be benecial for a number
of reasons. First, it could help infuse the site with a greater sense of community; not only are
169
people contributing to a large project (which no individual could possibly accomplish alone), but
users can also see how their photos t in with those of other users, which could help forge links
between users who may otherwise not notice a connection (oh, hes been to that cafe in Istanbul as
well!). Second, it would be very helpful in identifyingand encouraging people to ll ingaps
where people normally shoot few photos, i.e., the places in-between the famous sites. This could
become a scalable way of capturing a rich visual record of the entire world. Finally, the resulting
reconstruction could serve as a foundation onto which other types of information could be added:
labels on statues, buildings, paintings, and other objects; links to Wikipedia pages and other sources
of information; and comments and reviews of restaurants and other businesses. Creating such a
gathering place for the worlds images would not only require new algorithms for managing massive
amounts of data and predicting where new images should be taken, but also research into how to
best give people incentives to contribute. In spite of these challenges, I believe that this project could
have signicant impact, as a repository of data, a model of the world useful for education, research,
and virtual tourism, and a compelling social networking site.
170
BIBLIOGRAPHY
[1] Garmin GPS76 Owners Manual and Reference Guide. http://www8.garmin.com/
manuals/GPS76_OwnersManual.pdf.
[2] Aseem Agarwala, Maneesh Agrawala, Michael Cohen, David Salesin, and Richard Szeliski.
Photographing long scenes with multi-viewpoint panoramas. In SIGGRAPH Conf. Proc.,
pages 853861, 2006.
[3] Aseem Agarwala, Ke Colin Zheng, Chris Pal, Maneesh Agrawala, Michael Cohen, Brian
Curless, David Salesin, and Richard Szeliski. Panoramic video textures. In SIGGRAPH
Conf. Proc., pages 821827, 2005.
[4] Daniel G. Aliaga and Ingrid Carlbom. Plenoptic stitching: A scalable method for reconstruct-
ing 3D interactive walkthroughs. In SIGGRAPH Conf. Proc., pages 443450, 2001.
[5] Ingo Alth ofer, Gautam Das, David Dobkin, Deborah Joseph, and Jos e Soares. On sparse
spanners of weighted graphs. Discrete Comput. Geom., 9(1):81100, 1993.
[6] Steele Arbeeny and Deborah Silver. Spatial navigation of media streams. In Proc. Int. Conf.
on Multimedia, pages 467470, 2001.
[7] Sunil Arya, David M. Mount, Nathan S. Netanyahu, Ruth Silverman, and Angela Y. Wu. An
optimal algorithm for approximate nearest neighbor searching xed dimensions. J. of the
ACM, 45(6):891923, 1998.
[8] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features
(SURF). Computer Vision and Image Understanding, 110(3):346359, 2008.
[9] Benjamin B. Bederson. Photomesa: A zoomable image browser using quantum treemaps and
bubblemaps. In Proc. Symposium on User Interface Software and Technology, pages 7180,
2001.
[10] Pravin Bhat, C. Lawrence Zitnick, Noah Snavely, Aseem Agarwala, Maneesh Agrawala,
Brian Curless, Michael Cohen, and Sing Bing Kang. Using photographs to enhance videos
of a static scene. In Proc. Eurographics Symposium on Rendering, pages 327338, 2007.
[11] Olaf Booij, Zoran Zivkovic, and Ben Kr ose. Sparse appearance based modeling for robot
localization. In Proc. Int. Conf. on Intelligent Robots and Systems, pages 15101515, 2006.
171
[12] Fred L. Bookstein. Principal warps: Thin-plate splines and the decomposition of deforma-
tions. IEEE Trans. on Pattern Analysis and Machine Intelligence, 11(6):567585, 1989.
[13] Jean-Yves Bouguet. Camera calibration toolbox for Matlab. http://www.vision.
caltech.edu/bouguetj/calib_doc/index.html, 2001.
[14] Doug A. Bowman, David Koller, and Larry F. Hodges. Travel in immersive virtual envi-
ronments: An evaluation of viewpoint motion control techniques. In Proc. Virtual Reality
Annual Int. Symposium, pages 4552, 1997.
[15] Matthew Brown and David Lowe. Unsupervised 3D object recognition and reconstruction in
unordered datasets. In Proc. Int. Conf. on 3D Digital Imaging and Modeling, pages 5663,
2005.
[16] Aeron M. Buchanan and Andrew W. Fitzgibbon. Damped Newton algorithms for matrix
factorization with missing data. In Proc. IEEE Conf. on Computer Vision and Pattern Recog-
nition, pages 316322, 2005.
[17] Chris Buehler, Michael Bosse, Leonard McMillan, Steven Gortler, and Michael Cohen. Un-
structured lumigraph rendering. In SIGGRAPH Conf. Proc., pages 425432, 2001.
[18] Leizhen Cai. NP-completeness of minimum spanner problems. Discrete Appl. Math.,
48(2):187194, 1994.
[19] John Canny. A computational approach to edge detection. IEEE Trans. on Pattern Analysis
and Machine Intelligence, 8(6):679698, 1986.
[20] Shenchang Chen and Lance Williams. View interpolation for image synthesis. In SIGGRAPH
Conf. Proc., pages 279288, 1993.
[21] Shenchang Eric Chen. QuickTime VR An image-based approach to virtual environment
navigation. In SIGGRAPH Conf. Proc., pages 2938, 1995.
[22] L. Paul Chew. Constrained Delaunay triangulations. In Proc. Symposium on Computational
Geometry, pages 215222, 1987.
[23] Stephane Christy and Radu Horaud. Euclidean reconstruction: From paraperspective to per-
spective. In Proc. European Conf. on Computer Vision, pages 129140, 1996.
[24] Ondrej Chum, James Philbin, Josef Sivic, Michael Isard, and Andrew Zisserman. Total recall:
Automatic query expansion with a generative feature model for object retrieval. In Proc. Int.
Conf. on Computer Vision, 2007.
172
[25] Nico Cornelis, Kurt Cornelis, and Luc Van Gool. Fast compact city modeling for navigation
pre-visualization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition, pages
13391344, 2006.
[26] Cassidy Curtis and Chris Whitney. eyestilts: home of the telestereoscope. http://
eyestilts.com/.
[27] Rudolph P. Darken and John L. Sibert. Navigating large virtual spaces. Int. J. of Human-
Computer Interaction, 8(1):4971, 1996.
[28] Rudy P. Darken and John L. Sibert. A toolset for navigation in virtual environments. In Proc.
Symposium on User Interface Software and Technology, pages 157165, 1993.
[29] Marc Davis, Simon King, Nathan Good, and Risto Sarvas. From context to content: Lever-
aging context to infer media metadata. In Proc. Int. Conf. on Multimedia, pages 188195,
2004.
[30] Andrew J. Davison. Active search for real-time vision. In Proc. Int. Conf. on Computer
Vision, pages 6673, 2005.
[31] Marnix de Nijs. explodedviews. http://www.marnixdenijs.nl/exploded.htm.
[32] Paul E. Debevec, Camillo J. Taylor, and Jitendra Malik. Modeling and rendering architecture
from photographs: A hybrid geometry- and image-based approach. In SIGGRAPH Conf.
Proc., pages 1120, 1996.
[33] Digital photography review. http://www.dpreview.com/.
[34] Digital photography review glossary: Sensor sizes. http://www.dpreview.com/
learn/?/Glossary/Camera_System/sensor_sizes_01.%htm.
[35] E. W. Dijkstra. A note on two problems in connexion with graphs. Numerische Mathematik,
1(1):269271, December 1959.
[36] Cristina Russo dos Santos, Pascal Gros, Pierre Abel, Didier Loisel, Nicolas Trichaud, and
Jean-Pierre Paris. Metaphor-aware 3D navigation. In Proc. of the IEEE Symposium on Infor-
mation Vizualization, pages 155165. IEEE Computer Society, 2000.
[37] Steven M. Drucker, Curtis Wong, Asta Roseway, Steven Glenner, and Steven De Mar. Me-
diabrowser: Reclaiming the shoebox. In Proc. Working Conf. on Advanced Visual Interfaces,
pages 433436, 2004.
[38] Steven M. Drucker and David Zeltzer. Intelligent camera control in a virtual environment. In
Proc. of Graphics Interface, pages 190199, 1994.
173
[39] Chris Engels, Henrik Stewenius, and David Nist er. Bundle adjustment rules. In Proc. Sym-
posium on Photogrammetric Computer Vision, pages 266271, 2006.
[40] Boris Epshtein, Eyal Ofek, Yonathan Wexler, and Pusheng Zhang. Hierarchical photo or-
ganization using geometric relevance. In ACM Int. Symposium on Advances in Geographic
Information Systems, 2007.
[41] Everyscape. http://www.everyscape.com.
[42] How can you become a scape artist?. http://www.winsper.com/EveryScape/
scape_artist.php.
[43] Facebook. http://www.facebook.com.
[44] Steven Feiner, Blair MacIntyre, Tobias Hollerer, and Anthony Webster. A touring machine:
Prototyping 3D mobile augmented reality systems for exploring the urban environment. In
Proc. IEEE Int. Symposium on Wearable Computers, pages 7481, 1997.
[45] Martin Fischler and Robert Bolles. Random sample consensus: A paradigm for model t-
ting with applications to image analysis and automated cartography. Readings in Computer
Vision: Issues, Problems, Principles, and Paradigms, pages 726740, 1987.
[46] Andrew W. Fitzgibbon and Andrew Zisserman. Automatic camera recovery for closed or
open image sequences. In Proc. European Conf. on Computer Vision, pages 311326, 1998.
[47] Flickr. http://www.flickr.com.
[48] Michael L. Fredman and Robert Endre Tarjan. Fibonacci heaps and their uses in improved
network optimization algorithms. J. of the ACM, 34(3):596615, 1987.
[49] Yasutaka Furukawa and Jean Ponce. Accurate, dense, and robust multi-view stereopsis. In
Proc. IEEE Conf. on Computer Vision and Pattern Recognition, 2007.
[50] Tinsley A. Galyean. Guided navigation of virtual environments. In Proc. Symposium on
Interactive 3D Graphics, pages 103105, 1995.
[51] Gefos Geosystems. http://www.gefos.cz.
[52] James J. Gibson. The Ecological Approach to Visual Perception. Lawrence Erlbaum Asso-
ciates, 1979.
[53] Andreas Girgensohn, John Adcock, Matthew Cooper, Jonathon Foote, and Lynn Wilcox.
Simplifying the management of large photo collections. In Proc. Human-Computer Interac-
tion, pages 196203, 2003.
174
[54] Michael Goesele, Brian Curless, and Steven M. Seitz. Multi-view stereo revisited. In Proc.
IEEE Conf. on Computer Vision and Pattern Recognition, pages 24022409, 2006.
[55] Michael Goesele, Noah Snavely, Steven M. Seitz, Brian Curless, and Hugues Hoppe. Multi-
view stereo for community photo collections. In Proc. Int. Conf. on Computer Vision, 2007.
[56] Gene H. Golub and Charles F. Van Loan. Matrix Computations. The Johns Hopkins Univer-
sity Press, 1996.
[57] Google Earth. http://earth.google.com.
[58] Google Maps. http://maps.google.com.
[59] Steven J. Gortler, Radek Grzeszczuk, Richard Szeliski, and Michael F. Cohen. The lumi-
graph. In SIGGRAPH Conf. Proc., pages 4354, 1996.
[60] Graphviz - graph visualization software. http://www.graphviz.org/.
[61] Sudipto Guha and Samir Khuller. Approximation algorithms for connected dominating sets.
Algorithmica, 20(4):374387, 1998.
[62] Gabriele Guidi, Bernard Frischer, and Ignazio Lucenti. Rome Reborn - Virtualizing the
ancient imperial Rome. In Workshop on 3D Virtual Reconstruction and Visualization of
Complex Architectures, 2007.
[63] Chris Hand. A survey of 3D interaction techniques. Computer Graphics Forum, 16(5):269
281, 1997.
[64] Andrew J. Hanson and Eric A. Wernert. Constrained 3D navigation with 2D controllers. In
Proc. Conf. on Visualization, pages 175183, 1997.
[65] Frank Harary. Graph theory. Addison-Wesley, 1969.
[66] Richard I. Hartley. In defense of the eight-point algorithm. IEEE Trans. on Pattern Analysis
and Machine Intelligence, 19(6):580593, 1997.
[67] Richard I. Hartley and Frederik Schaffalitzky. PowerFactorization: 3D reconstruction with
missing or uncertain data. In Australia-Japan Advanced Workshop on Computer Vision, 2003.
[68] Richard I. Hartley and Andrew Zisserman. Multiple View Geometry. Cambridge University
Press, Cambridge, UK, 2004.
[69] Berthold K.P. Horn, Hugh M. Hilden, and Shahriar Negahdaripour. Closed-form solution of
absolute orientation using orthonormal matrices. J. Opt. Soc. of America A, 5:11271135,
July 1988.
175
[70] Intel math kernel library. http://www.intel.com/software/products/mkl.
[71] iPhoto. http://www.apple.com/ilife/iphoto.
[72] Alexander Jaffe, Mor Naaman, Tamir Tassa, and Marc Davis. Generating summaries and
visualization for large collections of geo-referenced photographs. In Proc. ACM SIGMM Int.
Workshop on Multimedia Information Retrieval, pages 8998, 2006.
[73] Yushi Jing and Shumeet Baluja. VisualRank: Applying PageRank to large-scale image
search. IEEE Trans. on Pattern Analysis and Machine Intelligence, 30(11):18771890, 2008.
[74] Rieko Kadobayashi and Katsumi Tanaka. 3D viewpoint-based photo search and information
browsing. In Proc. ACM Int. Conf. on Research and Development in Information Retrieval,
pages 621622, 2005.
[75] Hyunmo Kang and Ben Shneiderman. Visualization methods for personal photo collections:
Browsing and searching in the PhotoFinder. In Int. Conf. on Multimedia, pages 15391542,
2000.
[76] Sing Bing Kang, Peter-Pike Sloan, and Steven M. Seitz. Visual tunnel analysis for visibil-
ity prediction and camera planning. In Proc. IEEE Conf. on Computer Vision and Pattern
Recognition, pages 21952202, 2000.
[77] Lyndon Kennedy, Mor Naaman, Shane Ahern, Rahul Nair, and Tye Rattenbury. How Flickr
helps us make sense of the world: Context and content in community-contributed media
collections. In Proc. Int. Conf. on Multimedia, 2007.
[78] Azam Khan, Ben Komalo, Jos Stam, George Fitzmaurice, and Gordon Kurtenbach. Hov-
erCam: Interactive 3D navigation for proximal object inspection. In Proc. Symposium on
Interactive 3D Graphics and Games, pages 7380, 2005.
[79] Johannes Kopf, Matthew Uyttendaele, Oliver Deussen, and Michael F. Cohen. Capturing and
viewing gigapixel images. In SIGGRAPH Conf. Proc., 2007.
[80] Andreas Krause and Carlos Guestrin. Near-optimal observation selection using submodular
functions. In AAAI Conf. on Artical Intelligence, pages 16501654, 2007.
[81] E. Kruppa. Zur ermittlung eines objectes aus zwei perspektiven mit innerer orientierung.
Sitz.-Ber. Akad. Wiss., Wien, Math. Naturw. Kl., Abt. Ila., 122:19391948, 1913.
[82] Allan Kuchinsky, Celine Pering, Michael L. Creech, Dennis Freeze, Bill Serra, and Jacek
Gwizdka. FotoFile: A consumer multimedia organization and retrieval system. In Proc.
Conf. on Human Factors in Computing Systems, pages 496503, 1999.
176
[83] John Lasseter. Principles of traditional animation applied to 3D computer animation. In
SIGGRAPH Conf. Proc., pages 3544, 1987.
[84] Marc Levoy and Pat Hanrahan. Light eld rendering. In SIGGRAPH Conf. Proc., pages
3142, 1996.
[85] Marc Levoy, Kari Pulli, Brian Curless, Szymon Rusinkiewicz, David Koller, Lucas Pereira,
Matt Ginzton, Sean Anderson, James Davis, Jeremy Ginsberg, Jonathan Shade, and Duane
Fulk. The Digital Michelangelo Project: 3D scanning of large statues. In SIGGRAPH Conf.
Proc., pages 131144, 2000.
[86] Hongdong Li and Richard I. Hartley. Five-point motion estimation made easy. In Proc. Int.
Conf. on Pattern Recognition, pages 630633, 2006.
[87] AndrewLippman. Movie maps: An application of the optical videodisc to computer graphics.
In SIGGRAPH Conf. Proc., pages 3243, 1980.
[88] Live Search Maps. http://maps.live.com.
[89] Manolis Lourakis and Antonis Argyros. The design and implementation of a generic sparse
bundle adjustment software package based on the Levenberg-Marquardt algorithm. Technical
Report 340, Inst. of Computer Science-FORTH, Heraklion, Greece, 2004.
[90] David Lowe. Distinctive image features from scale-invariant keypoints. Int. J. of Computer
Vision, 60(2):91110, 2004.
[91] Daniel Martinec and Tomas Pajdla. Robust rotation and translation estimation in multiview
reconstruction. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition, 2007.
[92] Jiri Matas, Ondrej Chum, Urban Martin, and Tom as Pajdla. Robust wide baseline stereo from
maximally stable extremal regions. In Proc. of the British Machine Vision Conf., volume 1,
pages 384393, 2002.
[93] Neil McCurdy and William Griswold. A systems architecture for ubiquitous video. In Proc.
Int. Conf. on Mobile Systems, Applications, and Services, pages 114, 2005.
[94] Philip F. Mclauchlan. A batch/recursive algorithm for 3D scene reconstruction. In Proc.
IEEE Conf. on Computer Vision and Pattern Recognition, pages 738743, 2000.
[95] Leonard McMillan and Gary Bishop. Plenoptic modeling: An image-based rendering system.
In SIGGRAPH Conf. Proc., pages 3946, 1995.
[96] Eric N. Mortensen and William A. Barrett. Intelligent scissors for image composition. In
SIGGRAPH Conf. Proc., pages 191198, 1995.
177
[97] Mor Naaman, Andreas Paepcke, and Hector Garcia-Molina. From where to what: Metadata
sharing for digital photographs with geogr aphic coordinates. In Proc. Int. Conf. on Cooper-
ative Information Systems, pages 196217, 2003.
[98] Mor Naaman, Yee Jiun Song, Andreas Paepcke, and Hector Garcia-Molina. Automatic orga-
nization for digital photographs with geographic coordinates. In Proc. ACM/IEEE-CS Joint
Conf. on Digital Libraries, pages 5362, 2004.
[99] Michael Naimark. Aspen the verb: Musings on heritage and virtuality. Presence Journal,
15(3), June 2006.
[100] Kai Ni, Drew Steedly, and Frank Dellaert. Out-of-core bundle adjustment for large-scale 3D
reconstruction. In Proc. Int. Conf. on Computer Vision, 2007.
[101] David Nist er. Reconstruction fromuncalibrated sequences with a hierarchy of trifocal tensors.
In Proc. European Conf. on Computer Vision, pages 649663, 2000.
[102] David Nist er. An efcient solution to the ve-point relative pose problem. IEEE Trans.
Pattern Anal. Mach. Intell., 26(6):756777, 2004.
[103] David Nist er. Preemptive RANSAC for live structure and motion estimation. Mach. Vis.
Appl., 16(5):321329, 2005.
[104] David Nist er and Henrik Stew enius. Scalable recognition with a vocabulary tree. In Proc.
IEEE Conf. on Computer Vision and Pattern Recognition, pages 21182125, 2006.
[105] Jorge Nocedal and Stephen J. Wright. Numerical Optimization. Springer Series in Operations
Research. Springer-Verlag, New York, NY, 1999.
[106] Lawrence Page, Sergey Brin, Rajeev Motwani, and Terry Winograd. The PageRank citation
ranking: Bringing order to the Web. In Proc. Int. World Wide Web Conf., pages 161172,
1998.
[107] Finding paths through the worlds photos website. http://phototour.cs.
washington.edu/findingpaths/.
[108] Photo tourism website. http://phototour.cs.washington.edu/.
[109] Photosynth. http://photosynth.net/.
[110] The Photosynth photography guide. http://http://mslabs-777.vo.llnwd.
net/e1/documentation/Photosynth\%20Gu%ide\%20v7.pdf.
[111] Picasa. http://www.picasa.com.
178
[112] Picasa web albums. http://picasaweb.google.com.
[113] John C. Platt, Mary Czerwinski, and Brent A. Field. PhotoTOC: Automatic clustering for
browsing personal photographs. In IEEE Pacic Rim Conf, on Multimedia, 2003.
[114] Conrad J. Poelman and Takeo Kanade. A paraperspective factorization method for shape and
motion recovery. IEEE Trans. on Pattern Analysis and Machine Intelligence, 19(3):206218,
1997.
[115] Marc Pollefeys, David Nist er, Jan-Michael Frahm, Amir Akbarzadeh, Phillipos Mordohai,
Brian Clipp, Christopher Engels, David Gallup, S.-J. Kim, Paul Merrell, Christina Salmi,
Sudipta Sinha, Brad Talton, Liang Wang, Qing-Xiong Yang, Henrik Stew enius, Ruigang
Yang, Greg Welch, and Herman Towles. Detailed real-time urban 3D reconstruction from
video. Int. J. of Computer Vision, 78(2-3):143167, 2008.
[116] Marc Pollefeys, Luc van Gool, Maarten Vergauwen, Frank Verbiest, Kurt Cornelis, Jan Tops,
and Reinhard Koch. Visual modeling with a hand-held camera. Int. J. of Computer Vision,
59(3):207232, 2004.
[117] Suporn Pongnumkul, Jue Wang, and Michael Cohen. Creating map-based storyboards for
browsing tour videos. In Proc. Symposium on User Interface Software and Technology, pages
1322, 2008.
[118] Till Quack, Bastian Leibe, and Luc Van Gool. World-scale mining of objects and events
from community photo collections. In Proc. Int. Conf. on Content-based Image and Video
Retrieval, pages 4756, 2008.
[119] Jason Repko and Marc Pollefeys. 3D models from extended uncalibrated video sequences:
Addressing key-frame selection and projective drift. In Proc. Int. Conf. on 3-D Digital Imag-
ing and Modeling, pages 150157, 2005.
[120] D. P. Robertson and Roberto Cipolla. Building architectural models from many views using
map constraints. In Proc. European Conf. on Computer Vision, volume II, pages 155169,
2002.
[121] Roy A. Ruddle, Stephen J. Payne, and Dylan M. Jones. Navigating buildings in desk-top
virtual environments: Experimental investigations using extended navigational experience. J.
of Experimental Psychology: Applied, 3(2):143159, 1997.
[122] Alberto Ruiz, Pedro E. L opez de teruel, and Gin es Garca-mateos. A note on principal point
estimability. In Proc. Int. Conf. on Pattern Recognition, pages 1115, 2002.
[123] Bryan C. Russell, Antonio Torralba, Kevin P. Murphy, and William T. Freeman. LabelMe: A
database and web-based tool for image annotation. Int. J. of Computer Vision, 77(1-3):157
173, 2008.
179
[124] Frederik Schaffalitzky and Andrew Zisserman. Multi-view matching for unordered image
sets, or How do I organize my holiday snaps?. In Proc. European Conf. on Computer
Vision, volume 1, pages 414431, 2002.
[125] Grant Schindler, Frank Dellaert, and Sing Bing Kang. Inferring temporal order of images
from 3D structure. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition, 2007.
[126] Cordelia Schmid and Andrew Zisserman. Automatic line matching across views. In Proc.
IEEE Conf. on Computer Vision and Pattern Recognition, pages 666671, 1997.
[127] Seadragon. http://labs.live.com/Seadragon.aspx.
[128] Steven Seitz, Brian Curless, James Diebel, Daniel Scharstein, and Richard Szeliski. A com-
parison and evaluation of multi-view stereo reconstruction algorithms. In Proc. IEEE Conf.
on Computer Vision and Pattern Recognition, pages 519528, 2006.
[129] Ben Shneiderman and Hyunmo Kang. Direct annotation: A drag-and-drop strategy for label-
ing photos. In Proc. Int. Conf. on Information Visualisation, pages 8895, 2000.
[130] Ian Simon, Noah Snavely, , and Steven M. Seitz. Scene summarization for online image
collections. In Proc. Int. Conf. on Computer Vision, 2007.
[131] Karan Singh and Ravin Balakrishnan. Visualizing 3D scenes using non-linear projections
and data mining of previous camera movements. In Proc. Int. Conf. on Computer graphics,
virtual reality, visualisation and interaction in Africa, pages 4148, 2004.
[132] Josef Sivic, Biliana Kaneva, Antonio Torralba, Shai Avidan, and William T. Freeman. Cre-
ating and exploring a large photorealistic virtual space. In Proc. of the First IEEE Workshop
on Internet Vision, 2008.
[133] Josef Sivic and Andrew Zisserman. Video Google: A text retrieval approach to object match-
ing in video s. In Proc. Int. Conf. on Computer Vision, pages 14701477, 2003.
[134] Randall C. Smith and Peter Cheeseman. On the representation and estimation of spatial
uncertainly. Int. J. Robotics Research, 5(4):5668, 1987.
[135] Smugmug photo sharing. http://www.smugmug.com.
[136] Noah Snavely, Rahul Garg, Steven M. Seitz, and Richard Szeliski. Finding paths through the
worlds photos. In SIGGRAPH Conf. Proc., 2008.
[137] Noah Snavely, Steven M. Seitz, and Richard Szeliski. Photo tourism: Exploring photo col-
lections in 3D. In SIGGRAPH Conf. Proc., pages 835846, 2006.
180
[138] Diomidis D. Spinellis. Position-annotated photographs: A geotemporal web. IEEE Pervasive
Computing, 2(2):7279, 2003.
[139] Drew Steedly, Irfan Essa, and Frank Delleart. Spectral partitioning for structure from motion.
In Proc. Int. Conf. on Computer Vision, pages 996103, 2003.
[140] Drew Steedly and Irfan A. Essa. Propagation of innovative information in non-linear least-
squares structure from motion. In Proc. Int. Conf. on Computer Vision, pages 223229, 2001.
[141] Peter Sturm and Bill Triggs. A factorization based algorithm for multi-image projective
structure and motion. In Proc. European Conf. on Computer Vision, pages 709720, 1996.
[142] Ivan E. Sutherland. A head-mounted three dimensional display. In Proc. Fall Joint Computer
Conf., pages 757764, 1968.
[143] R. Szeliski and S. B. Kang. Recovering 3D shape and motion from image streams using
nonlinear least squares. J. of Visual Communication and Image Representation, 5(1):1028,
March 1994.
[144] Richard Szeliski. Image alignment and stitching: A tutorial. Foundations and Trends in
Computer Graphics and Computer Vision, 2(1), December 2006.
[145] Desney S. Tan, George G. Robertson, and Mary Czerwinski. Exploring 3D navigation: Com-
bining speed-coupled ying with orbiting. In Proc. Conf. on Human Factors in Computing
Systems, pages 418425. ACM Press, 2001.
[146] Hiroya Tanaka, Masatoshi Arikawa, and Ryosuke Shibasaki. A 3-D photo collage system for
spatial navigations. In Revised Papers from the Second Kyoto Workshop on Digital Cities II,
Computational and Sociological Approaches, pages 305316, 2002.
[147] Camillo J. Taylor. VideoPlus: A method for capturing the structure and appearance of immer-
sive environments. IEEE Transactions on Visualization and Computer Graphics, 8(2):171
182, April-June 2002.
[148] Carlo Tomasi and Takeo Kanade. Shape and motion from image streams under orthography:
A factorization method. Int. J. of Computer Vision, 9(2):137154, 1992.
[149] Kentaro Toyama, Ron Logan, and Asta Roseway. Geographic location tags on digital images.
In Proc. Int. Conf. on Multimedia, pages 156166, 2003.
[150] Bill Triggs, Phil McLauchlan, Richard I. Hartley, and Andrew Fitzgibbon. Bundle adjustment
a modern synthesis. In Vision Algorithms: Theory and Practice, volume 1883 of Lecture
Notes in Computer Science, pages 298372, 2000.
181
[151] U.S. Geological Survey. http://www.usgs.com.
[152] Matthew Uyttendaele, Antonio Criminisi, Sing Bing Kang, Simon Winder, Richard Szeliski,
and Richard I. Hartley. Image-based interactive exploration of real-world environments. IEEE
Computer Graphics and Applications, 24(3):5263, 2004.
[153] Maarten Vergauwen and Luc Van Gool. Web-based 3D reconstruction service. Mach. Vis.
Appl., 17(2):321329, 2006.
[154] Luc Vincent. Taking online maps down to street level. IEEE Computer, 40(12):118120,
December 2007.
[155] Luis von Ahn and Laura Dabbish. Labeling images with a computer game. In Proc. Conf. on
Human Factors in Computing Systems, pages 319326, 2004.
[156] Matthias Wandel. Exif jpeg header manipulation tool. http://www.sentex.net/
mwandel/jhead/.
[157] Colin Ware and Steven Osborne. Exploration and virtual camera control in virtual three
dimensional environments. In Proc. Symposium on Interactive 3D Graphics, pages 175183.
ACM Press, 1990.
[158] Wikipedia. http://www.wikipedia.org.
[159] Windows Live Local - Virtual Earth Technology Preview. http://preview.local.
live.com.
[160] Zhengyou Zhang. A exible new technique for camera calibration. IEEE Trans. on Pattern
Analysis and Machine Intelligence, 22(11):13301334, 2000.
182
Appendix A
ESTIMATING THE FOCAL LENGTH OF A DIGITAL PHOTO FROM EXIF
TAGS
Almost every digital camera manufactured in the last few years embeds useful data about each
captured photo into the JPEG les that get stored to the cameras memory. This information, en-
coded in Exif (exchangeable image le format) tags, often includes exposure time, focus, aperture,
whether the ash was activated, and focal length. The last of these, focal length, is especially useful
for structure from motion. This appendix describes how to extract the focal length from the Exif
tags of a digital photo and to convert it to the pixel units that can be directly used in structure from
motion. The following pieces of information are required for this conversion:
1. The focal length estimate (in mm), f
mm
.
2. The width of the camera CCD (in mm), CCD
mm
.
3. The dimensions of the image (in pixels), w
pixels
, h
pixels
.
Once all of these numbers are known, computing the focal length in pixel, f
pixels
, can be done via a
simple unit conversion:
f
pixels
=
f
mm
CCD
mm
(maxw
pixels
, h
pixels
)
This computation assumes that the image has not been cropped or warped in any way (aside from
simple resizing of the image). For instance, if the focal length of a photo is listed in the Exif tags as
5.4mm, we used, say, a Canon PowerShot S100 (with a CCD width of 5.27mm) to take the photo,
and the image resolution is 1600 1200, then
f
pixels
=
5.4
5.27
1600 = 1639.47.
To extract f
mm
from the Exif tags, any Exif tag reader will do. JHead [156] is a good free one
which works from the command line. The resolution of an image is also embedded in the Exif tags.
183
The camera manufacturer and model name are also usually embedded as well, and for many camera
models CCD
mm
can be found online on camera review sites such as Digital Photography Review
[33]. For instance, the specications page for the Canon PowerShot S100 on Digital Photography
Review lists the dimensions of the sensor as 1/2.7 (5.27 x 3.96 mm). Some cameras embed the
CCD width directly into the Exif tags, but in my experience, this number is often unreliable.
As with the example above, sometimes the CCD size is given as a ratio of inches (e.g., 1/2.7).
These sizes stem from standard diameters of television camera tubes, and the conversion from these
units to millimeters is not straightforward; Digital Photo Review describes the conversion in its
glossary [34].
184
Appendix B
COMPLEXITY ANALYSIS OF INCREMENTAL STRUCTURE FROM MOTION
The running time of the incremental structure from motion (SfM) algorithm of Chapter 3 de-
pends on the exact schedule in which the images are added to the reconstruction, and is dominated
by the calls to SBA [89] after every batch of new images is reconstructed. Let S(t) be the running
time of a call to SBA when there are t images currently in the reconstruction. Let T(n, r) be the total
amount of time spent inside SBA during SfM given n input images, r of which are added during
each iteration of SfM.
1
Recall that S(t) = (t
3
), so there exists N such that, for all t > N,
c
1
t
3
S(t) c
2
t
3
(B.1)
for some constants c
1
and c
2
.
Case 1: r =
n
k
. First, suppose that
n
k
images are added per iteration of SfM, i.e., we are consider-
ing T
_
n,
n
k
_
. We have that
T
_
n,
n
k
_
=
k

i=1
S
_
i
n
k
_
.
For
n
k
> N, the bounds in Equation B.1 apply. Considering just the upper bound, we have that:
T
_
n,
n
k
_

k

i=1
c
2
_
i
n
k
_
3

i=1
c
2
_
k
n
k
_
3
= c
2
kn
3
.
1
In practice, the number of images added will vary between iterations. For simplicity of analysis, I only consider the
case where the same number of images are added each time.
185
It follows that T
_
n,
n
k
_
= O(n
3
). A lower bound proportional to n
3
can also be shown, thus
T
_
n,
n
k
_
= (n
3
).
Case 2: r = . Now let us consider the case when a constant number of images are added during
each iteration, for a total of
n

iterations (assuming divides n). We have that

T(n, ) =
n/

i=1
S(i ). (B.2)
Let us group the rst
N

terms in the summation in Equation B.2 into a single constant, C (assuming,

without loss of generality, that divides N); the remaining terms (where i > N) can then be
bounded using Equation B.1. Considering just the upper bound gives:
T(n, ) =
n/

i=1
S(i ) (B.3)
=
N/

i=1
S(i ) +
n/

i=N/+1
S(i ) (B.4)
= C +
n/

i=N/+1
S(i ) (B.5)
C +
n/

i=N/+1
c
2
(i )
3
(B.6)
= C +c
1

3
n/

i=N/+1
i
3
(B.7)
= C +c
1

3
_
_
n/

i=1
i
3

N/

i=1
i
3
_
_
. (B.8)
186
The sum over i
3
for 1 i N/ is another constant, which we denote by D:
T(n, ) C +c
1

3
_
_
n/

i=1
i
3
D
_
_
(B.9)
=
_
C c
1

3
D
_
+c
1

3
n/

i=1
i
3
(B.10)
=
_
C c
1

3
D
_
+c
1

3
_
_
n

_ _
n

+ 1
_
2
_
2
(B.11)
= O(n
4
). (B.12)
Therefore T(n, ) = O(n
4
). Using the lower bound in Eq. B.1, we can also show that T(n, ) =
(n
4
) through a very similar analysis, replacing
T(n, ) C +
n/

i=N/+1
c
2
(i )
3
in Eq. B.6 with
T(n, ) C +
n/

i=N/+1
c
1
(i )
3
Hence, T(n, ) = (n
4
).
187
Appendix C
LINE SEGMENT RECONSTRUCTION
Once the point cloud has been reconstructed and the camera positions recovered using structure
from motion (SfM), the SfM pipeline reconstructs 3D line segments in the scene. Line segments can
be useful for generating better triangulated morphs (Chapter 5, Section 5.1.2) and for enhancing the
renderings of architectural scenes. Line segment reconstruction from images has been investigated
in work such as [126]. My 3Dline segment reconstruction algorithmis similar, and has the following
steps:
1. Detect 2D line segments in each image.
2. Derive a set of candidate 2D line segment matches by comparing line segments between pairs
of nearby images.
3. Find sets of mutually consistent matches across multiple images (line tracks) above a certain
size and triangulate the matching 2D line segments to obtain a 3D line segment.
These steps are now described in more detail.
To detect 2D line segments, I use Canny edge detection [19], followed by an edge-linking step
to form edge chains. These chains may have any shape, and may be closed; the next step is to break
the chains into sub-chains that approximate line segments. To do so, for each edge chain, I t a line
to the chain using orthogonal regression. If the furthest point on the chain from the tted line has a
distance to the line greater than a threshold, the chain is broken in two at that extremal point, then
this procedure is recursively applied two the two new chains. Finally, I remove chains smaller than
a threshold, and store the two endpoints of the line segment approximating of each of the remaining
chains. I use S(I) to denote the set of 2D line segments found in image I.
I then match line segments between images. For each image I
j
, I rst compute a set of other
images with which to match. This set should contain images whose cameras are close to camera C
j
and looking in roughly the same direction. To compute this set, I nd the 32 cameras closest to C
j
188
and remove those whose viewing directions are at an angle greater than a threshold to the viewing
direction of C
j
.
Next, for each camera C
k
in the set, I consider each line segment s S(I
j
). Each line segment
t S(I
k
) is labeled as a candidate match of s if t meets the following two conditions:
1. The endpoints of t are not too far away from the epipolar lines of the endpoints of s in image
I
k
.
2. The L
2
distance between a strip of intensity values around s and a strip of intensity values
around t is not too large. The intensity values are sampled along epipolar lines, and each strip
is normalized for bias and gain before their L
2
distance is computed. After normalization, I
use a threshold of 0.3 to reject candidate matches.
After computing candidate matches between pairs of neighbors, I compute connected compo-
nents of candidate matches, as was done with SIFT feature matches in the SfM pipeline (Chapter 3,
Section 3.1). Unlike with keypoint matching, however, I now know where each photograph was
taken, so we can immediately check connected components for geometric consistency. For each
connected component of matches with a subset of consistent line segments of size at least four, I
create a 3D line segment. The 3D line segment is created by triangulating corresponding endpoints
of the 2D line segments; the resulting 3D points form the endpoints of the 3D line segment.
189
Appendix D
PHOTO CREDITS
I would like to thank the following people for generously allowing me to use and reproduce
their photos in my work. Photos from these Flickr users were used in the Photo Tourism project,
and appear in Chapter 5 or the associated video.
Holly Ables
http://www.flickr.com/photos/tmlens/
of Nashville, TN
Rakesh Agrawal http://www.flickr.com/photos/asmythie/
Pedro Alcocer http://www.flickr.com/photos/pealco/
Julien Avarre http://www.flickr.com/photos/eole/
Rael Bennett http://www.flickr.com/photos/spooky05/
Loc Bernard http://www.flickr.com/photos/loic1967/
Nicole Bratt http://www.flickr.com/photos/nicolebratt/
Nicholas Brown http://www.flickr.com/photos/nsgbrown/
Domenico Calojero
1
http://www.flickr.com/photos/mikuzz/
DeGanta Choudhury http://www.flickr.com/photos/deganta/
dan clegg http://www.flickr.com/photos/mathmandan/
Claude Covo-Farchi http://www.flickr.com/photos/bip/
Alper C u gun http://www.flickr.com/photos/alper/
W. Garth Davis http://www.flickr.com/photos/garth/
Stamatia Eliakis http://www.flickr.com/photos/12537899@N00/
Dawn Endico
2
http://www.flickr.com/photos/candiedwomanire/
Silvana M. Felix http://www.flickr.com/photos/jadoreparis/
Jeroen Hamers http://www.flickr.com/photos/jhamers/
Caroline H ardter http://www.flickr.com/photos/dotpolka/
Mary Harrsch http://www.flickr.com/photos/44124324682@N01/
Molly Hazelton http://www.flickr.com/photos/mollx/
Bill Jennings
3
http://www.flickr.com/photos/mrjennings/
Michelle Joo http://www.flickr.com/photos/maz/
Tommy Keswick http://www.flickr.com/photos/mrwilloby/
Kirsten Gilbert Krenicky http://www.flickr.com/photos/kirsten gilbert krenicky/
Giampaolo Macorig http://www.flickr.com/photos/60405541@N00/
Erin K Malone
4
http://www.flickr.com/photos/erinmalone/
Daryoush Mansouri http://www.flickr.com/photos/daryoush/
1
mikuzz@gmail.com
2
endico@gmail.com
3
Supported by grants from the National Endowment for the Humanities and the Fund for Teachers.
4
photographs copyright 2005
190
Paul Meidinger http://www.flickr.com/photos/pmeidinger/
Laurete de
http://www.flickr.com/photos/21guilherme/
Albuquerque Mouazan
Callie Neylan http://www.flickr.com/photos/williamcatherine/
Robert Norman http://www.flickr.com/photos/bobtribe/
Dirk Olbertz http://www.flickr.com/photos/dirkolbertz/
Dave Ortman http://www.flickr.com/photos/dortman/
George Owens http://www.flickr.com/photos/georgeowens/
Claire Elizabeth Poulin http://www.flickr.com/photos/claireep/
David R. Preston http://www.flickr.com/photos/dosbears/
Jim Sellers and Laura Kluver http://www.flickr.com/photos/jim and laura/
Peter Snowling http://www.flickr.com/photos/19654387@N00/
Rom Srinivasan http://www.flickr.com/photos/romsrini/
Jeff Allen Wallen Photography http://www.flickr.com/photos/goondockjeff/
Daniel West http://www.flickr.com/photos/curious/
Todd A. Van Zandt http://www.flickr.com/photos/vanzandt/
Dario Zappal` a http://www.flickr.com/photos/zappa/
Susan Elnadi http://www.flickr.com/photos/30596986@N00/
I also would like to acknowledge the following people whose photographs were reproduced for
Chapter 5 under Creative Commons licenses:
Shoshanah http://www.flickr.com/photos/shoshanah
1
Dan Kamminga http://www.flickr.com/photos/dankamminga
1
Tjeerd Wiersma http://www.flickr.com/photos/tjeerd
1
Manogamo http://www.flickr.com/photos/se-a-vida-e
2
Ted Wang http://www.flickr.com/photos/mtwang
3
Arnet http://www.flickr.com/photos/gurvan
3
Rebekah Martin http://www.flickr.com/photos/rebekah
3
Jean Ruaud http://www.flickr.com/photos/jrparis
3
Imran Ali http://www.flickr.com/photos/imran
3
Scott Goldblatt http://www.flickr.com/photos/goldblatt
3
Todd Martin http://www.flickr.com/photos/tmartin
4
Steven http://www.flickr.com/photos/graye
4
ceriess http://www.flickr.com/photos/ceriess
1
Cory Pi na http://www.flickr.com/photos/corypina
3
mark gallagher http://www.flickr.com/photos/markgallagher
1
Celia http://www.flickr.com/photos/100n30th
3
Carlo B. http://www.flickr.com/photos/brodo
3
Kurt Naks http://www.flickr.com/photos/kurtnaks
4
Anthony M. http://www.flickr.com/photos/antmoose
1
Virginia G http://www.flickr.com/photos/vgasull
3
1
http://creativecommons.org/licenses/by/2.0/
2
http://creativecommons.org/licenses/by-nd/2.0/
3
http://creativecommons.org/licenses/by-nc-nd/2.0/
4
http://creativecommons.org/licenses/by-nc/2.0/
191
Collection credit and copyright notice for Moon and Half Dome, 1960, by Ansel Adams: Collection
Center for Creative Photography, University of Arizona, c _Trustees of The Ansel Adams Publishing
Rights Trust.
The following people allowed me to use photos for the Pathnder project. Their photos appear
in Chapter 6 and in the accompanying video.
Storm Crypt http://www.flickr.com/photos/storm-crypt
James McPherson http://www.flickr.com/photos/jamesontheweb
Wolfgang Wedenig http://www.flickr.com/photos/wuschl2202
AJP79 http://www.flickr.com/photos/90523335@N00
Tony Thompson http://www.flickr.com/photos/14489588@N00
Warren Buckley http://www.flickr.com/photos/studio85
Keith Barlow http://www.flickr.com/photos/keithbarlow
beautifulcataya http://www.flickr.com/photos/beautifulcataya
Smiley Apple http://www.flickr.com/photos/smileyapple
crewealexandra http://www.flickr.com/photos/28062159@N00
Ian Turk http://www.flickr.com/photos/ianturk
Randy Fish http://www.flickr.com/photos/randyfish
Justin Kauk http://www.flickr.com/photos/justinkauk
Airplane Lane http://www.flickr.com/photos/photons
Katie Holmes http://www.flickr.com/photos/katieholmes
Cher Kian Tan http://www.flickr.com/photos/70573485@N00
Erin Longdo http://www.flickr.com/photos/eel
James McKenzie http://www.flickr.com/photos/jmckenzie
Eli Garrett http://www.flickr.com/photos/portenaeli
Francesco Gasparetti http://www.flickr.com/photos/gaspa
Emily Galopin http://www.flickr.com/photos/theshrtone
Sandro Mancuso http://www.flickr.com/photos/worldwalker
Ian Monroe http://www.flickr.com/photos/eean
Noam Freedman http://www.flickr.com/photos/noamf
morbin http://www.flickr.com/photos/morbin
Margrethe Store http://www.flickr.com/photos/margrethe
Eugenia and Julian http://www.flickr.com/photos/eugeniayjulian
Allyson Boggess http://www.flickr.com/photos/allysonkalea
Ed Costello http://www.flickr.com/photos/epc
Paul Kim http://www.flickr.com/photos/fmg2001
Susan Elnadi http://www.flickr.com/people/30596986@N00
Mathieu Pinet http://www.flickr.com/photos/altermativ
c _ Ariane Gaudefroy http://www.flickr.com/photos/kicouette
Briana Baldwin http://www.flickr.com/photos/breezy421
Andrew Nguyen http://www.flickr.com/photos/nguy0833
Curtis Townson http://www.flickr.com/photos/fifty50
Rob Thatcher
http://www.flickr.com/photos/pondskater
(rob@hypereal.co.uk)
Greg Scher http://www.flickr.com/photos/gregscher
192
VITA
Keith (Noah) Snavely grew up in Tucson, Arizona, and received his B.S. in Computer Science
and Math fromthe University of Arizona in 2003. He then joined the Ph.D. programof the Computer
Science and Engineering department at the University of Washington, where he worked with Steven
M. Seitz and Richard Szeliski. In 2008, he received the Doctor of Philosophy degree.